Add files using upload-large-folder tool

DeepSeek-VL2/deepseek_vl2/models/__init__.py

@@ -0,0 +1,26 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+from .processing_deepseek_vl_v2 import DeepseekVLV2Processor
+from .modeling_deepseek_vl_v2 import DeepseekVLV2ForCausalLM
+
+__all__ = [
+    "DeepseekVLV2Processor",
+    "DeepseekVLV2ForCausalLM",
+]

DeepSeek-VL2/deepseek_vl2/models/configuration_deepseek.py

@@ -0,0 +1,210 @@
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
+class DeepseekV2Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`DeepseekV2Model`]. It is used to instantiate an DeepSeek
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the DeepSeek-V2 with multi-latent attention.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 102400):
+            Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`DeepseekV2Model`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 11008):
+            Dimension of the MLP representations.
+        moe_intermediate_size (`int`, *optional*, defaults to 1407):
+            Dimension of the MoE representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        n_shared_experts (`int`, *optional*, defaults to None):
+            Number of shared experts, None means dense model.
+        n_routed_experts (`int`, *optional*, defaults to None):
+            Number of routed experts, None means dense model.
+        routed_scaling_factor (`float`, *optional*, defaults to 1.0):
+            Scaling factor or routed experts.
+        topk_method (`str`, *optional*, defaults to `gready`):
+            Topk method used in routed gate.
+        n_group (`int`, *optional*, defaults to None):
+            Number of groups for routed experts.
+        topk_group (`int`, *optional*, defaults to None):
+            Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups).
+        num_experts_per_tok (`int`, *optional*, defaults to None):
+            Number of selected experts, None means dense model.
+        moe_layer_freq (`int`, *optional*, defaults to 1):
+            The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers.
+        first_k_dense_replace (`int`, *optional*, defaults to 0):
+            Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).
+                                                            \--k dense layers--/
+        norm_topk_prob (`bool`, *optional*, defaults to False):
+            Whether to normalize the weights of the routed experts.
+        scoring_func (`str`, *optional*, defaults to 'softmax'):
+            Method of computing expert weights.
+        aux_loss_alpha (`float`, *optional*, defaults to 0.001):
+            Auxiliary loss weight coefficient.
+        seq_aux = (`bool`, *optional*, defaults to True):
+            Whether to compute the auxiliary loss for each individual sample.
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 2048):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            End of stream token id.
+        pretraining_tp (`int`, *optional*, defaults to 1):
+            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
+            document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is
+            necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
+            issue](https://github.com/pytorch/pytorch/issues/76232).
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling
+            strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
+            `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
+            `max_position_embeddings` to the expected new maximum.
+        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        use_mla (`bool`, *optional*, defaults to `True`): Use multi-latent attention or multi-head attention. If True,
+            the model will use multi-latent attention, otherwise, it will use multi-head attention.
+
+    ```python
+    >>> from transformers import DeepseekV2Model, DeepseekV2Config
+
+    >>> # Initializing a Deepseek-V2 style configuration
+    >>> configuration = DeepseekV2Config()
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "deepseek_v2"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=102400,
+        hidden_size=4096,
+        intermediate_size=11008,
+        moe_intermediate_size = 1407,
+        num_hidden_layers=30,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        n_shared_experts = None,
+        n_routed_experts = None,
+        ep_size = 1,
+        routed_scaling_factor = 1.0,
+        kv_lora_rank = 512,
+        q_lora_rank = 1536,
+        qk_rope_head_dim = 64,
+        v_head_dim = 128,
+        qk_nope_head_dim = 128,
+        topk_method = 'gready',
+        n_group = None,
+        topk_group = None,
+        num_experts_per_tok = None,
+        moe_layer_freq = 1,
+        first_k_dense_replace = 0,
+        norm_topk_prob = False,
+        scoring_func = 'softmax',
+        aux_loss_alpha = 0.001,
+        seq_aux = True,
+        hidden_act="silu",
+        max_position_embeddings=2048,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=100000,
+        eos_token_id=100001,
+        pretraining_tp=1,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        use_mla=True,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.moe_intermediate_size = moe_intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.n_shared_experts = n_shared_experts
+        self.n_routed_experts = n_routed_experts
+        self.ep_size = ep_size
+        self.routed_scaling_factor = routed_scaling_factor
+        self.kv_lora_rank = kv_lora_rank
+        self.q_lora_rank = q_lora_rank
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.topk_method = topk_method
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.num_experts_per_tok = num_experts_per_tok
+        self.moe_layer_freq = moe_layer_freq
+        self.first_k_dense_replace = first_k_dense_replace
+        self.norm_topk_prob = norm_topk_prob
+        self.scoring_func = scoring_func
+        self.aux_loss_alpha = aux_loss_alpha
+        self.seq_aux = seq_aux
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = float(rms_norm_eps)
+        self.pretraining_tp = pretraining_tp
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.use_mla = use_mla
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

DeepSeek-VL2/deepseek_vl2/models/conversation.py

@@ -0,0 +1,310 @@
+"""
+From https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
+"""
+
+import dataclasses
+from enum import IntEnum, auto
+from typing import Any, Dict, List
+
+
+class SeparatorStyle(IntEnum):
+    """Separator styles."""
+
+    DeepSeek = auto()
+    DeepSeekV2 = auto()
+    PLAIN = auto()
+    ALIGNMENT = auto()
+
+
+@dataclasses.dataclass
+class Conversation:
+    """A class that manages prompt templates and keeps all conversation history."""
+
+    # The name of this template
+    name: str
+    # The template of the system prompt
+    system_template: str = "{system_message}"
+    # The system message
+    system_message: str = ""
+    # The names of two roles
+    roles: List[str] = (("USER", "ASSISTANT"),)
+    # All messages. Each item is (role, message).
+    messages: List[List[str]] = ()
+    # The number of few shot examples
+    offset: int = 0
+    # The separator style and configurations
+    sep_style: SeparatorStyle = SeparatorStyle.DeepSeek
+    sep: str = "\n"
+    sep2: str = None
+    # Stop criteria (the default one is EOS token)
+    stop_str: str = None
+    # Stops generation if meeting any token in this list
+    stop_token_ids: List[int] = None
+
+    def get_prompt(self) -> str:
+        """Get the prompt for generation."""
+        system_prompt = self.system_template.format(system_message=self.system_message)
+        if self.sep_style == SeparatorStyle.DeepSeek:
+            seps = [self.sep, self.sep2]
+            if system_prompt == "" or system_prompt is None:
+                ret = ""
+            else:
+                ret = system_prompt + seps[0]
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    ret += role + ": " + message + seps[i % 2]
+                else:
+                    ret += role + ":"
+            return ret
+        elif self.sep_style == SeparatorStyle.DeepSeekV2:
+            seps = [self.sep, self.sep2]
+            if system_prompt == "" or system_prompt is None:
+                ret = ""
+            else:
+                ret = system_prompt + seps[0]
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    if role == "User":
+                        ret += "<｜sft▁begin｜>\n" + message + self.sep #<｜sft▁begin｜>User Input<｜sft▁end｜>\nResponse<｜end▁of▁sentence｜>
+                    else:
+                        ret += message + self.sep2
+                else:
+                    ret = ret
+            return ret
+
+        elif self.sep_style == SeparatorStyle.PLAIN:
+            seps = [self.sep, self.sep2]
+            ret = ""
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    if i % 2 == 0:
+                        ret += message + seps[i % 2]
+                    else:
+                        ret += message + seps[i % 2]
+                else:
+                    ret += ""
+            return ret
+        elif self.sep_style == SeparatorStyle.ALIGNMENT:
+            seps = [self.sep, self.sep2]
+            ret = ""
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    if i % 2 == 0:
+                        ret += '<image>\n' + seps[i % 2]
+                    else:
+                        ret += message + seps[i % 2]
+                else:
+                    ret += ""
+            return ret
+        else:
+            raise ValueError(f"Invalid style: {self.sep_style}")
+
+    def set_system_message(self, system_message: str):
+        """Set the system message."""
+        self.system_message = system_message
+
+    def append_message(self, role: str, message: str):
+        """Append a new message."""
+        self.messages.append([role, message])
+
+    def update_last_message(self, message: str):
+        """Update the last output.
+
+        The last message is typically set to be None when constructing the prompt,
+        so we need to update it in-place after getting the response from a model.
+        """
+        self.messages[-1][1] = message
+
+    def reset_message(self):
+        """Reset a new message."""
+        self.messages = []
+
+    def to_gradio_chatbot(self):
+        """Convert the conversation to gradio chatbot format."""
+        ret = []
+        for i, (role, msg) in enumerate(self.messages[self.offset :]):
+            if i % 2 == 0:
+                ret.append([msg, None])
+            else:
+                ret[-1][-1] = msg
+        return ret
+
+    def to_openai_api_messages(self):
+        """Convert the conversation to OpenAI chat completion format."""
+        system_prompt = self.system_template.format(system_message=self.system_message)
+        ret = [{"role": "system", "content": system_prompt}]
+
+        for i, (_, msg) in enumerate(self.messages[self.offset :]):
+            if i % 2 == 0:
+                ret.append({"role": "user", "content": msg})
+            else:
+                if msg is not None:
+                    ret.append({"role": "assistant", "content": msg})
+        return ret
+
+    def copy(self):
+        return Conversation(
+            name=self.name,
+            system_template=self.system_template,
+            system_message=self.system_message,
+            roles=self.roles,
+            messages=[[x, y] for x, y in self.messages],
+            offset=self.offset,
+            sep_style=self.sep_style,
+            sep=self.sep,
+            sep2=self.sep2,
+            stop_str=self.stop_str,
+            stop_token_ids=self.stop_token_ids,
+        )
+
+    def dict(self):
+        return {
+            "template_name": self.name,
+            "system_message": self.system_message,
+            "roles": self.roles,
+            "messages": self.messages,
+            "offset": self.offset,
+        }
+
+
+# A global registry for all conversation templates
+conv_templates: Dict[str, Conversation] = {}
+
+
+def register_conv_template(template: Conversation, override: bool = False):
+    """Register a new conversation template."""
+    if not override:
+        assert template.name not in conv_templates, f"{template.name} has been registered."
+
+    conv_templates[template.name] = template
+
+
+def get_conv_template(name: str) -> Conversation:
+    """Get a conversation template."""
+    return conv_templates[name].copy()
+
+
+# register_conv_template(
+#     Conversation(
+#         name="deepseek",
+#         system_template="{system_message}",
+#         # system_message="You are a helpful assistant. Please answer truthfully and write out your "
+#         # "thinking step by step to be sure you get the right answer.",
+#         system_message="",
+#         roles=("User", "Assistant"),
+#         messages=(),
+#         offset=0,
+#         sep_style=SeparatorStyle.DeepSeek,
+#         sep="\n\n",
+#         sep2="<｜end▁of▁sentence｜>",
+#         stop_token_ids=[100001],
+#         stop_str=["User:", "<｜end▁of▁sentence｜>"]
+#     )
+# )
+register_conv_template(
+    Conversation(
+        name="deepseek",
+        system_template="{system_message}",
+        # system_message="You are a helpful assistant. Please answer truthfully and write out your "
+        # "thinking step by step to be sure you get the right answer.",
+        system_message="",
+        roles=("<|User|>", "<|Assistant|>"),
+        messages=(),
+        offset=0,
+        sep_style=SeparatorStyle.DeepSeek,
+        sep="\n\n",
+        sep2="<｜end▁of▁sentence｜>",
+        stop_token_ids=[100001],
+        stop_str=["User:", "<｜end▁of▁sentence｜>"]
+    )
+)
+# register_conv_template(
+#     Conversation(
+#         name="deepseekv2",
+#         system_template="{system_message}",
+#         system_message="",
+#         roles=("User", "Assistant"),
+#         messages=(),
+#         offset=0,
+#         sep_style=SeparatorStyle.DeepSeekV2,
+#         sep="\n<｜sft▁end｜>",
+#         sep2="<｜end▁of▁sentence｜>",
+#         stop_token_ids=[100001],
+#         stop_str=["User:", "<｜end▁of▁sentence｜>"]
+#     )
+# )
+register_conv_template(
+    Conversation(
+        name="deepseekv2",
+        system_template="{system_message}",
+        system_message="",
+        roles=("|<User>|", "|<Assistant>|"),
+        messages=(),
+        offset=0,
+        sep_style=SeparatorStyle.DeepSeekV2,
+        sep="\n<｜sft▁end｜>",
+        sep2="<｜end▁of▁sentence｜>",
+        stop_token_ids=[100001],
+        stop_str=["User:", "<｜end▁of▁sentence｜>"]
+    )
+)
+
+
+register_conv_template(
+    Conversation(
+        name="plain",
+        system_template="",
+        system_message="",
+        roles=("", ""),
+        messages=(),
+        offset=0,
+        sep_style=SeparatorStyle.PLAIN,
+        sep="",
+        sep2="",
+        stop_token_ids=[100001],
+        stop_str=['</s>'],
+    )
+)
+
+
+register_conv_template(
+    Conversation(
+        name="alignment",
+        system_template="",
+        system_message="",
+        roles=("", ""),
+        messages=(),
+        offset=0,
+        sep_style=SeparatorStyle.ALIGNMENT,
+        sep="",
+        sep2="",
+        stop_token_ids=[100001],
+        stop_str=['</s>'],
+    )
+)
+
+
+if __name__ == "__main__":
+    print("deepseek template:")
+    conv = get_conv_template("deepseek")
+    conv.append_message(conv.roles[0], "Hello!")
+    conv.append_message(conv.roles[1], "Hi! This is Tony.")
+    conv.append_message(conv.roles[0], "Who are you?")
+    conv.append_message(conv.roles[1], "I am a helpful assistant.")
+    conv.append_message(conv.roles[0], "How are you?")
+    conv.append_message(conv.roles[1], None)
+    print(conv.get_prompt())
+
+    print("deepseekv2 template:")
+    conv = get_conv_template("deepseekv2")
+    conv.append_message(conv.roles[0], "Hello!")
+    conv.append_message(conv.roles[1], "Hi! This is Tony.")
+    conv.append_message(conv.roles[0], "Who are you?")
+    conv.append_message(conv.roles[1], "I am a helpful assistant.")
+    conv.append_message(conv.roles[0], "How are you?")
+    conv.append_message(conv.roles[1], None)
+    print(conv.get_prompt())

DeepSeek-VL2/deepseek_vl2/models/processing_deepseek_vl_v2.py

@@ -0,0 +1,675 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+from dataclasses import dataclass
+from typing import Dict, Tuple, List, Literal, Optional
+import math
+
+import torch
+from torch.nn.utils.rnn import pad_sequence
+import torchvision.transforms as T
+from transformers import LlamaTokenizerFast
+from transformers.processing_utils import ProcessorMixin
+from PIL import Image, ImageOps
+
+from .conversation import get_conv_template
+
+
+def select_best_resolution(image_size, candidate_resolutions):
+    # used for cropping
+    original_width, original_height = image_size
+    best_fit = None
+    max_effective_resolution = 0
+    min_wasted_resolution = float("inf")
+
+    for width, height in candidate_resolutions:
+        scale = min(width / original_width, height / original_height)
+        downscaled_width, downscaled_height = int(original_width * scale), int(original_height * scale)
+        effective_resolution = min(downscaled_width * downscaled_height, original_width * original_height)
+        wasted_resolution = (width * height) - effective_resolution
+
+        if effective_resolution > max_effective_resolution or (effective_resolution == max_effective_resolution and wasted_resolution < min_wasted_resolution):
+            max_effective_resolution = effective_resolution
+            min_wasted_resolution = wasted_resolution
+            best_fit = (width, height)
+
+    return best_fit
+
+
+class DictOutput(object):
+    def keys(self):
+        return self.__dict__.keys()
+
+    def __getitem__(self, item):
+        return self.__dict__[item]
+
+    def __setitem__(self, key, value):
+        self.__dict__[key] = value
+
+
+# 对于inference sample也可以维护input_ids，反正最后不会用到
+@dataclass
+class VLChatProcessorOutput(DictOutput):
+    sft_format: str
+    input_ids: torch.LongTensor
+    target_ids: torch.LongTensor
+    images: torch.Tensor
+    images_seq_mask: torch.BoolTensor
+    images_spatial_crop: torch.LongTensor
+    num_image_tokens: List[int]
+
+    def __len__(self):
+        return len(self.input_ids)
+
+
+@dataclass
+class BatchCollateOutput(DictOutput):
+    sft_format: List[str]
+    input_ids: torch.LongTensor
+    labels: torch.LongTensor
+    images: torch.Tensor
+    attention_mask: torch.Tensor
+    images_seq_mask: torch.BoolTensor
+    images_spatial_crop: torch.LongTensor
+    seq_lens: List[int]
+
+    def to(self, device, dtype=torch.bfloat16):
+        self.input_ids = self.input_ids.to(device)
+        self.labels = self.labels.to(device)
+        self.attention_mask = self.attention_mask.to(device)
+        self.images_seq_mask = self.images_seq_mask.to(device)
+        self.images_spatial_crop = self.images_spatial_crop.to(device)
+        self.images = self.images.to(device=device, dtype=dtype)
+        return self
+
+
+class ImageTransform(object):
+    def __init__(
+            self,
+            mean: Optional[Tuple[float, float, float]] = (0.5, 0.5, 0.5),
+            std: Optional[Tuple[float, float, float]] = (0.5, 0.5, 0.5),
+            normalize: bool = True
+    ):
+        self.mean = mean
+        self.std = std
+        self.normalize = normalize
+
+        transform_pipelines = [
+            T.ToTensor()
+        ]
+
+        if normalize:
+            transform_pipelines.append(T.Normalize(mean, std))
+
+        self.transform = T.Compose(transform_pipelines)
+
+    def __call__(self, pil_img: Image.Image):
+        x = self.transform(pil_img)
+        return x
+
+
+
+class DeepseekVLV2Processor(ProcessorMixin):
+    tokenizer_class = ("LlamaTokenizer", "LlamaTokenizerFast")
+    attributes = ["tokenizer"]
+
+    def __init__(
+            self,
+            tokenizer: LlamaTokenizerFast,
+            candidate_resolutions: Tuple[Tuple[int, int]],
+            patch_size: int,
+            downsample_ratio: int,
+            image_mean: Tuple[float, float, float] = (0.5, 0.5, 0.5),
+            image_std: Tuple[float, float, float] = (0.5, 0.5, 0.5),
+            normalize: bool = True,
+            image_token: str = "<image>",
+            pad_token: str = "<｜▁pad▁｜>",
+            add_special_token: bool = False,
+            sft_format: str = "deepseek",
+            mask_prompt: bool = True,
+            ignore_id: int = -100,
+            **kwargs,
+    ):
+
+        self.candidate_resolutions = candidate_resolutions
+        self.image_size = candidate_resolutions[0][0]
+        self.patch_size = patch_size
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.normalize = normalize
+        self.downsample_ratio = downsample_ratio
+
+        self.image_transform = ImageTransform(mean=image_mean, std=image_std, normalize=normalize)
+        self.tokenizer = tokenizer
+        self.tokenizer.padding_side = 'left'  # must set this，padding side with make a difference in batch inference
+
+        # add the pad_token as special token to use 'tokenizer.pad_token' and 'tokenizer.pad_token_id'
+        if tokenizer.pad_token is None:
+            self.tokenizer.add_special_tokens({'pad_token': pad_token})
+        print(f"Add pad token = ['{pad_token}'] to the tokenizer\n"
+              f"{pad_token}:{tokenizer.encode(pad_token, add_special_tokens=False)[0]}")
+
+        # add image token
+        image_token_id = self.tokenizer.vocab.get(image_token)
+        if image_token_id is None:
+            special_tokens = [image_token]
+            special_tokens_dict = {"additional_special_tokens": special_tokens}
+            self.tokenizer.add_special_tokens(special_tokens_dict)
+        self.image_token_id = self.tokenizer.vocab.get(image_token)
+        print(f"Add image token = ['{image_token}'] to the tokenizer\n"
+              f"{image_token}:{tokenizer.encode(image_token, add_special_tokens=False)[0]}")
+
+        # add five special tokens for grounding-related tasks
+        # <|ref|>, <|/ref|>, <|det|>, <|/det|>, <|grounding|>
+        special_tokens = ['<|ref|>', '<|/ref|>', '<|det|>', '<|/det|>', '<|grounding|>']
+        special_tokens_dict = {"additional_special_tokens": special_tokens}
+        self.tokenizer.add_special_tokens(special_tokens_dict)
+        print(f"Add grounding-related tokens = {special_tokens} to the tokenizer with input_ids\n"
+              f"<|ref|>:{tokenizer.encode('<|ref|>', add_special_tokens=False)[0]}\n"
+              f"<|/ref|>:{tokenizer.encode('<|/ref|>', add_special_tokens=False)[0]}\n"
+              f"<|det|>:{tokenizer.encode('<|det|>', add_special_tokens=False)[0]}\n"
+              f"<|/det|>:{tokenizer.encode('<|/det|>', add_special_tokens=False)[0]}\n"
+              f"<|grounding|>:{tokenizer.encode('<|grounding|>', add_special_tokens=False)[0]}")
+
+        # add special tokens for SFT data
+        special_tokens = ["<|User|>", "<|Assistant|>"]
+        special_tokens_dict = {"additional_special_tokens": special_tokens}
+        self.tokenizer.add_special_tokens(special_tokens_dict)
+        print(f"Add chat tokens = {special_tokens} to the tokenizer with input_ids\n"
+              f"<|User|>:{tokenizer.encode('<|User|>', add_special_tokens=False)[0]}\n"
+              f"<|Assistant|>:{tokenizer.encode('<|Assistant|>', add_special_tokens=False)[0]}\n")
+
+        self.image_token = image_token
+        self.pad_token = pad_token
+        self.add_special_token = add_special_token
+        self.sft_format = sft_format
+        self.mask_prompt = mask_prompt
+        self.ignore_id = ignore_id
+
+        super().__init__(
+            tokenizer,
+            **kwargs,
+        )
+
+    def new_chat_template(self):
+        conv = get_conv_template(self.sft_format)
+        return conv
+
+    def format_messages(
+            self,
+            conversations: List[Dict[str, str]],
+            sft_format: str = "deepseek",
+            system_prompt: str = "",
+    ):
+        """
+        Applies the SFT template to conversation.
+
+        Args:
+            conversations (List[Dict]): A List of messages.
+            sft_format (str, optional): The format of the SFT template to use. Defaults to "deepseek".
+            system_prompt (str, optional): The system prompt to use in the SFT template. Defaults to "".
+
+        Returns:
+            sft_prompt (str): The formatted text.
+        """
+
+        conv = get_conv_template(sft_format)
+        conv.set_system_message(system_prompt)
+        for message in conversations:
+            conv.append_message(message["role"], message["content"].strip())
+        sft_prompt = conv.get_prompt().strip()
+
+        return sft_prompt
+
+    def format_messages_v2(self, messages, pil_images, systems=None):
+        """play the role of format_messages_v2 and get_images_info in the last version"""
+        tokenized_data = []
+        masked_tokenized_data = []  # labels
+        images_list = []
+        images_seq_mask = []
+        images_spatial_crop = []
+        num_image_tokens = []
+
+        image_index = 0
+
+        conv = get_conv_template(self.sft_format)
+        conv_system_message = conv.system_message
+
+        for idx, message in enumerate(messages):
+            if idx == 0:
+                tokenized_data += [self.bos_id]
+                masked_tokenized_data += [self.bos_id]
+                images_seq_mask += [False]
+                conv.system_message = conv_system_message
+            else:
+                conv.system_message = ''
+
+            if message['role'] == conv.roles[0] or message['role'] == "user":
+                conv.reset_message()
+                conv.append_message(conv.roles[0], str(message['content']).strip())
+                conv.append_message(conv.roles[1], '')
+                formatted_question = conv.get_prompt()
+                tokenized_str, images, seq_mask, spatial_crop, n_image_tokens = self.tokenize_with_images(
+                    formatted_question,
+                    pil_images[image_index: image_index + formatted_question.count(self.image_token)],
+                    bos=False,
+                    eos=False,
+                    cropping=len(pil_images) <= 2
+                )
+                image_index += formatted_question.count(self.image_token)
+
+                tokenized_data += tokenized_str
+                if self.mask_prompt:
+                    masked_tokenized_data += [self.ignore_id] * len(tokenized_str)
+                else:
+                    masked_tokenized_data += tokenized_str
+                images_list += images
+                images_seq_mask += seq_mask
+                images_spatial_crop += spatial_crop
+                num_image_tokens += n_image_tokens
+
+            elif message['role'] == conv.roles[1] or message['role'] == "assistant":
+                formatted_answer = message['content'].strip()
+                assert formatted_answer.count(
+                    self.image_token) == 0, f"there should be no {self.image_token} in the assistant's reply, but got {messages}"
+                tokenized_str, images, seq_mask, spatial_crop, n_image_tokens = self.tokenize_with_images(
+                    formatted_answer,
+                    [],
+                    bos=False,
+                    eos=True,
+                    cropping=len(pil_images) <= 2)
+
+                tokenized_data += tokenized_str
+                masked_tokenized_data += tokenized_str
+                images_seq_mask += seq_mask
+
+            elif message['role'] == 'system' or message['role'] == 'deepseekapi-sys':
+                # 如果message里面有system，那就只允许出现在message的第一句，同时conv原本的system就会失效
+                assert idx == 0, 'system information should only exist in the begining of the conversation'
+                formatted_system = message['content'].strip()
+                tokenized_str = self.encode(formatted_system, bos=False, eos=False)
+                tokenized_data += tokenized_str
+                if self.mask_prompt:
+                    masked_tokenized_data += [self.ignore_id] * len(tokenized_str)
+                else:
+                    masked_tokenized_data += tokenized_str
+                seq_mask = [False] * len(tokenized_str)
+                images_seq_mask += seq_mask
+
+            else:
+                assert False, f"Unknown role: {message['role']}"
+
+        assert len(tokenized_data) == len(
+            images_seq_mask), f"format_messages_v2: tokenized_str's length {len(tokenized_str)} is not equal to imags_seq_mask's length {len(images_seq_mask)}"
+        assert len(images_spatial_crop) == len(num_image_tokens), f"image number should be compatible"
+
+        return tokenized_data, masked_tokenized_data, images_list, images_seq_mask, images_spatial_crop, num_image_tokens
+
+    def format_prompts(
+            self,
+            prompts: str,
+            sft_format: str = "deepseek",
+            system_prompt: str = "",
+    ):
+        """
+        Applies the SFT template to prompts.
+
+        Args:
+            prompts (str): the non-sft formatted prompt;
+            sft_format (str, optional): The format of the SFT template to use. Defaults to "deepseek".
+            system_prompt (str, optional): The system prompt to use in the SFT template. Defaults to "".
+
+        Returns:
+            sft_prompt (str): The formatted text.
+        """
+
+        conv = get_conv_template(sft_format)
+        conv.set_system_message(system_prompt)
+        conv.append_message(conv.roles[0], prompts.strip())
+        conv.append_message(conv.roles[1], "")
+
+        sft_prompt = conv.get_prompt().strip()
+
+        return sft_prompt
+
+    @property
+    def bos_id(self):
+        return self.tokenizer.bos_token_id
+
+    @property
+    def eos_id(self):
+        return self.tokenizer.eos_token_id
+
+    @property
+    def pad_id(self):
+        return self.tokenizer.pad_token_id
+
+    def encode(self, text: str, bos: bool = True, eos: bool = False):
+        t = self.tokenizer.encode(text, add_special_tokens=False)
+
+        if bos:
+            t = [self.bos_id] + t
+        if eos:
+            t = t + [self.eos_id]
+
+        return t
+
+    def decode(self, t: List[int], **kwargs) -> str:
+        return self.tokenizer.decode(t, **kwargs)
+
+    def process_one(
+            self,
+            prompt: str = None,
+            conversations: List[Dict[str, str]] = None,
+            images: List[Image.Image] = None,
+            apply_sft_format: bool = False,
+            inference_mode: bool = True,
+            system_prompt: str = "",
+            **kwargs,
+    ):
+        """
+
+        Args:
+            prompt (str): the formatted prompt;
+            conversations (List[Dict]): conversations with a list of messages;
+            images (List[ImageType]): the list of images;
+            apply_sft_format (bool): if prompt is not None, then apply the SFT format to prompt;
+                if conversations is not None, then it will always apply the SFT format to conversations;
+            inference_mode (bool): if True, then remove the last eos token;
+            system_prompt (str): the system prompt;
+            **kwargs:
+
+        Returns:
+            outputs (BaseProcessorOutput): the output of the processor,
+                - input_ids (torch.LongTensor): [N + image tokens]
+                - target_ids (torch.LongTensor): [N + image tokens]
+                - images (torch.FloatTensor): [n_images, 3, H, W]
+                - image_id (int): the id of the image token
+                - num_image_tokens (List[int]): the number of image tokens
+        """
+
+        assert (
+                prompt is None or conversations is None
+        ), "prompt and conversations cannot be used at the same time."
+
+        if prompt is None:
+            # apply sft format
+            sft_format = self.format_messages(
+                conversations=conversations,
+                sft_format=self.sft_format,
+                system_prompt=system_prompt,
+            )
+            tokenized_str, masked_tokenized_str, images_list, images_seq_mask, images_spatial_crop, num_image_tokens = self.format_messages_v2(
+                conversations, images)
+        else:
+            if apply_sft_format:
+                sft_format = self.format_prompts(
+                    prompts=prompt,
+                    sft_format=self.sft_format,
+                    system_prompt=system_prompt
+                )
+            else:
+                sft_format = prompt
+            tokenized_str, images_list, images_seq_mask, images_spatial_crop, num_image_tokens = self.tokenize_with_images(
+                sft_format, images, bos=True, eos=True, cropping=len(images) <= 2)
+            masked_tokenized_str = []
+            for token_index in tokenized_str:
+                if token_index != self.image_token_id:
+                    masked_tokenized_str.append(token_index)
+                else:
+                    masked_tokenized_str.append(self.ignore_id)
+
+        assert len(tokenized_str) == len(images_seq_mask) == len(masked_tokenized_str), \
+            (f"tokenized_str's length {len(tokenized_str)}, input_ids' length {len(masked_tokenized_str)}, "
+             f"imags_seq_mask's length {len(images_seq_mask)}, are not equal")
+
+        input_ids = torch.LongTensor(tokenized_str)
+        target_ids = torch.LongTensor(masked_tokenized_str)
+        images_seq_mask = torch.tensor(images_seq_mask, dtype=torch.bool)
+
+        # set input_ids < 0 | input_ids == self.image_token_id as ignore_id
+        target_ids[(input_ids < 0) | (input_ids == self.image_token_id)] = self.ignore_id
+        input_ids[input_ids < 0] = self.pad_id
+
+        if inference_mode:
+            # 去掉结尾的eos token
+            assert input_ids[-1] == self.eos_id
+            input_ids = input_ids[:-1]
+            target_ids = target_ids[:-1]
+            images_seq_mask = images_seq_mask[:-1]
+
+        if len(images_list) == 0:
+            images = torch.zeros((1, 3, self.image_size, self.image_size))
+            images_spatial_crop = torch.zeros((1, 2), dtype=torch.long)
+        else:
+            images = torch.stack(images_list, dim=0)
+            images_spatial_crop = torch.tensor(images_spatial_crop, dtype=torch.long)
+
+        prepare = VLChatProcessorOutput(
+            sft_format=sft_format,
+            input_ids=input_ids,
+            target_ids=target_ids,
+            images=images,
+            images_seq_mask=images_seq_mask,
+            images_spatial_crop=images_spatial_crop,
+            num_image_tokens=num_image_tokens
+        )
+
+        return prepare
+
+    def __call__(
+            self,
+            *,
+            prompt: str = None,
+            conversations: List[Dict[str, str]] = None,
+            images: List[Image.Image] = None,
+            apply_sft_format: bool = False,
+            force_batchify: bool = True,
+            inference_mode: bool = True,
+            system_prompt: str = "",
+            **kwargs,
+    ):
+        """
+
+        Args:
+            prompt (str): the formatted prompt;
+            conversations (List[Dict]): conversations with a list of messages;
+            images (List[ImageType]): the list of images;
+            apply_sft_format (bool): if prompt is not None, then apply the SFT format to prompt;
+                if conversations is not None, then it will always apply the SFT format to conversations;
+            force_batchify (bool): force batchify the inputs;
+            inference_mode (bool): if True, then remove the last eos token;
+            system_prompt (str): the system prompt;
+            **kwargs:
+
+        Returns:
+            outputs (BaseProcessorOutput): the output of the processor,
+                - input_ids (torch.LongTensor): [N + image tokens]
+                - images (torch.FloatTensor): [n_images, 3, H, W]
+                - image_id (int): the id of the image token
+                - num_image_tokens (List[int]): the number of image tokens
+        """
+
+        prepare = self.process_one(
+            prompt=prompt,
+            conversations=conversations,
+            images=images,
+            apply_sft_format=apply_sft_format,
+            inference_mode=inference_mode,
+            system_prompt=system_prompt
+        )
+
+        if force_batchify:
+            prepare = self.batchify([prepare])
+
+        return prepare
+
+    def tokenize_with_images(
+            self,
+            conversation: str,
+            images: List[Image.Image],
+            bos: bool = True,
+            eos: bool = True,
+            cropping: bool = True,
+    ):
+        """Tokenize text with <image> tags."""
+        assert conversation.count(self.image_token) == len(images)
+        text_splits = conversation.split(self.image_token)
+        images_list, images_seq_mask, images_spatial_crop = [], [], []
+        num_image_tokens = []
+        tokenized_str = []
+        for text_sep, image in zip(text_splits, images):
+            """encode text_sep"""
+            tokenized_sep = self.encode(text_sep, bos=False, eos=False)
+            tokenized_str += tokenized_sep
+            images_seq_mask += [False] * len(tokenized_sep)
+
+            """select best resolution for anyres"""
+            if cropping:
+                best_width, best_height = select_best_resolution(image.size, self.candidate_resolutions)
+            else:
+                best_width, best_height = self.image_size, self.image_size
+            # print(image.size, (best_width, best_height)) # check the select_best_resolutions func
+
+            """process the global view"""
+            global_view = ImageOps.pad(image, (self.image_size, self.image_size),
+                                       color=tuple(int(x * 255) for x in self.image_transform.mean))
+            images_list.append(self.image_transform(global_view))
+
+            """process the local views"""
+            local_view = ImageOps.pad(image, (best_width, best_height),
+                                      color=tuple(int(x * 255) for x in self.image_transform.mean))
+            for i in range(0, best_height, self.image_size):
+                for j in range(0, best_width, self.image_size):
+                    images_list.append(
+                        self.image_transform(local_view.crop((j, i, j + self.image_size, i + self.image_size))))
+
+            """record height / width crop num"""
+            num_width_tiles, num_height_tiles = best_width // self.image_size, best_height // self.image_size
+            images_spatial_crop.append([num_width_tiles, num_height_tiles])
+
+            """add image tokens"""
+            h = w = math.ceil((self.image_size // self.patch_size) / self.downsample_ratio)
+            # global views tokens h * (w + 1), 1 is for line seperator
+            tokenized_image = [self.image_token_id] * h * (w + 1)
+            # add a seperator between global and local views
+            tokenized_image += [self.image_token_id]
+            # local views tokens, (num_height_tiles * h) * (num_width_tiles * w + 1)
+            tokenized_image += [self.image_token_id] * (num_height_tiles * h) * (num_width_tiles * w + 1)
+
+            tokenized_str += tokenized_image
+            images_seq_mask += [True] * len(tokenized_image)
+            num_image_tokens.append(len(tokenized_image))
+            # print(width_crop_num, height_crop_num, len(tokenized_image)) # test the correctness of the number of image-related tokens
+
+        """process the last text split"""
+        tokenized_sep = self.encode(text_splits[-1], bos=False, eos=False)
+        tokenized_str += tokenized_sep
+        images_seq_mask += [False] * len(tokenized_sep)
+
+        """add the bos and eos tokens"""
+        if bos:
+            tokenized_str = [self.bos_id] + tokenized_str
+            images_seq_mask = [False] + images_seq_mask
+        if eos:
+            tokenized_str = tokenized_str + [self.eos_id]
+            images_seq_mask = images_seq_mask + [False]
+
+        assert len(tokenized_str) == len(
+            images_seq_mask), f"tokenize_with_images func: tokenized_str's length {len(tokenized_str)} is not equal to imags_seq_mask's length {len(images_seq_mask)}"
+
+        return tokenized_str, images_list, images_seq_mask, images_spatial_crop, num_image_tokens
+
+    def batchify(
+            self,
+            sample_list: List[VLChatProcessorOutput],
+            padding: Literal["left", "right"] = "left"
+    ) -> BatchCollateOutput:
+        """
+        Preprocesses the inputs for multimodal inference.
+
+        Args:
+            sample_list (List[VLChatProcessorOutput]): A list of VLChatProcessorOutput.
+            padding (str): The padding method. Defaults to "left".
+
+        Returns:
+            BatchCollateOutput: A dictionary of the inputs to use for multimodal inference.
+        """
+
+        batched_sft_format = [sample.sft_format for sample in sample_list]
+        batched_input_ids = [sample.input_ids for sample in sample_list]
+        batched_labels = [sample.target_ids for sample in sample_list]
+        batched_images_seq_mask = [sample["images_seq_mask"] for sample in sample_list]
+        seq_lens = [len(sample) for sample in sample_list]
+
+        """padding input_ids and images_seq_mask"""
+        if padding == "left":
+            # the tokenizer is default to pad at left
+            ## TODO, You're using a LlamaTokenizerFast tokenizer.
+            #   Please note that with a fast tokenizer, using the `__call__` method is faster than
+            #   using a method to encode the text followed by a call to the `pad` method to get a padded encoding.
+            padded_input_ids = self.tokenizer.pad({"input_ids": batched_input_ids})
+            batched_input_ids, batched_attention_mask = padded_input_ids["input_ids"], padded_input_ids[
+                "attention_mask"].bool()
+            batched_labels = self.tokenizer.pad({"input_ids": batched_labels})["input_ids"]
+            batched_labels[batched_labels == self.pad_id] = self.ignore_id  # labels正常不会出现pad_id，无需额外保护
+            batched_images_seq_mask = self.tokenizer.pad({"input_ids": batched_images_seq_mask})["input_ids"]
+            batched_images_seq_mask[batched_images_seq_mask == self.pad_id] = False
+        else:
+            batched_input_ids = pad_sequence(batched_input_ids, batch_first=True, padding_value=self.pad_id)
+            batched_labels = pad_sequence(batched_labels, batch_first=True, padding_value=self.ignore_id)
+            batched_images_seq_mask = pad_sequence(batched_images_seq_mask, batch_first=True, padding_value=0)
+            batched_attention_mask = batched_input_ids != self.pad_id
+
+        """padding images to max_patch_num"""
+        max_n_patches = max(sample["images"].shape[0] for sample in sample_list)
+        batched_images = []
+        for sample in sample_list:
+            images = sample["images"]
+            n_pads = max_n_patches - images.shape[0]
+            if n_pads > 0:
+                pad_images = torch.zeros((n_pads, *images.shape[1:]), dtype=images.dtype)
+                images = torch.cat([images, pad_images], dim=0)
+            batched_images.append(images)
+        batched_images = torch.stack(batched_images, dim=0)
+
+        """padding images_spatial_crop to max_n_images"""
+        max_n_images = max(sample["images_spatial_crop"].shape[0] for sample in sample_list)
+        batched_images_spatial_crop = []
+        for sample in sample_list:
+            images_spatial_crop = sample["images_spatial_crop"]
+            n_pads = max_n_images - sample["images_spatial_crop"].shape[0]
+            if n_pads > 0:
+                pad_images_spatial_crop = torch.full((n_pads, 2), 0, dtype=images_spatial_crop.dtype)
+                images_spatial_crop = torch.cat([images_spatial_crop, pad_images_spatial_crop], dim=0)
+            batched_images_spatial_crop.append(images_spatial_crop)
+        batched_images_spatial_crop = torch.stack(batched_images_spatial_crop, dim=0)
+
+        batched_samples = BatchCollateOutput(
+            input_ids=batched_input_ids,
+            attention_mask=batched_attention_mask,
+            labels=batched_labels,
+            images=batched_images,
+            images_seq_mask=batched_images_seq_mask,
+            images_spatial_crop=batched_images_spatial_crop,
+            sft_format=batched_sft_format,
+            seq_lens=seq_lens
+        )
+
+        return batched_samples

EAGLE/README.md

@@ -0,0 +1,355 @@
+<p align="center" width="100%">
+<img src="assets/Logo.png"  width="80%" height="80%">
+</p>
+<!-- # EAGLE: Exploring The Design Space for Multimodal LLMs with Mixture of Encoders -->
+
+---
+
+[![Code License](https://img.shields.io/badge/Code%20License-Apache_2.0-green.svg)](https://github.com/tatsu-lab/stanford_alpaca/blob/main/LICENSE)
+[![Model License](https://img.shields.io/badge/MODEL%20License-CC%20By%20NC%204.0-red.svg)](MODEL_LICENSE)
+
+[[arXiv](https://arxiv.org/pdf/2408.15998)] [[HuggingFace](https://huggingface.co/papers/2408.15998)] [[Demo](https://huggingface.co/spaces/NVEagle/Eagle-X5-13B-Chat)] 
+[[Model Zoo](https://huggingface.co/NVEagle)] [[Data](https://huggingface.co/datasets/shi-labs/Eagle-1.8M)]
+
+
+## Introduction
+
+Eagle is a family of Vision-Centric High-Resolution Multimodal LLMs. It presents a thorough exploration to strengthen multimodal LLM perception with a mixture of vision encoders and different input resolutions. The model contains a channel-concatenation-based "CLIP+X" fusion for vision experts with different architectures (ViT/ConvNets) and knowledge (detection/segmentation/OCR/SSL). The resulting family of Eagle models support up to over 1K input resolution and obtain strong results on multimodal LLM benchmarks, especially resolution-sensitive tasks such as optical character recognition and document understanding.
+
+<div align="center">
+<img src="assets/fig-teaser.jpg" width="90%">
+</div>
+
+
+## Updates
+- [TODO] Vision encoder model weights with pre-alignment.
+- [2024/09] 🔥 Release models trained on the [Cambrian-1](https://huggingface.co/collections/nyu-visionx/cambrian-data-6667ce801e179b4fbe774e11) data.
+- [2024/09] Provide an [example](#evaluation) for evaluation.
+- [2024/08] Release the Eagle-X5-13B-Chat online [demo](https://huggingface.co/spaces/NVEagle/Eagle-X5-13B-Chat).
+- [2024/08] Release the [Eagle-SFT-1.8M](https://huggingface.co/datasets/shi-labs/Eagle-1.8m) data.
+- [2024/08] 🔥 Release models trained on the [LLaVA-1.5 Pre-train](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain) and [Eagle-SFT-1.8M](https://huggingface.co/datasets/shi-labs/Eagle-1.8m) data.
+- [2024/08] Release training and inference code of Eagle.
+- [2024/06] 🏆 Winning the 2nd Place in CVPR24 Challenge on Driving with Language [Website](https://opendrivelab.com/challenge2024/#driving_with_language).
+- [2024/05] Serving as the 2D VLM pre-training for [OmniDrive](https://github.com/NVlabs/OmniDrive).
+
+
+## Contents
+- [Models & Performance](#models--performance)
+- [Visual Examples](#visual-examples)
+- [Install](#install)
+- [Training Data](#training-data)
+- [Checkpoint Preparation](#checkpoint-preparation)
+- [Training](#training)
+- [Inference](#inference)
+- [Evaluation](#evaluation)
+- [Gradio Demo](#gradio-demo)
+
+
+## Models & Performance
+Models trained on the [LLaVA-1.5 Pre-train](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain) and [Eagle-SFT-1.8M](https://huggingface.co/datasets/shi-labs/Eagle-1.8m) data are available to download here.
+| Model&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; | LLM&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; | Pretrain&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; | SFT | GQA | MME | MMMU(Val) | OCR | SQA(I) | POPE | TextVQA | InfoVQA | VizWiz | SEED(I) | VQAv2 | MathVista | MMBench | ChartQA | DocVQA |
+|------|------|------|------|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|
+| [Eagle-X4-7B](https://huggingface.co/NVEagle/Eagle-X4-7B) | Vicuna-7B | LLaVA-v1.5 | 1.8M | 64.8 | 1561 | 34.9 | 540 | 70.5 | 88.4 | 70.9 | 47.4 | 50.8 | 73.4 | 83.4 | 37.3 | 67.8 | 67.5 | 78.8 |
+| [Eagle-X5-7B](https://huggingface.co/NVEagle/Eagle-X5-7B) | Vicuna-7B | LLaVA-v1.5 | 1.8M | 64.9 | 1528 | 36.3 | 529 | 69.8 | 88.8 | 71.2 | 47.4 | 54.4 | 73.9 | 83.4 | 37.0 | 68.4 | 67.8 | 78.6 |
+| [Eagle-X4-13B](https://huggingface.co/NVEagle/Eagle-X4-13B) | Vicuna-13B | LLaVA-v1.5 | 1.8M | 66.3 | 1627 | 36.9 | 561 | 73.1 | 87.7 | 73.9 | 50.7 | 56.2 | 74.4 | 83.8 | 37.6 | 69.9 | 70.5 | 79.9 |
+| [Eagle-X5-13B](https://huggingface.co/NVEagle/Eagle-X5-13B) | Vicuna-13B |  LLaVA-v1.5 | 1.8M | 66.2 | 1609 | 36.6 | 574 | 72.8 | 87.8 | 74.2 | 51.8 | 59.3 | 74.1 | 83.8 | 38.8 | 69.2 | 69.9 | 79.4 |
+
+Models trained on the [Cambrian-1](https://huggingface.co/collections/nyu-visionx/cambrian-data-6667ce801e179b4fbe774e11) data are available to download here.
+|      |      | Knowledge |      |      |      |      | General |      |      |      |      | Document |      |      |      |      | Vision |      |      |
+|------|------|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|:----:|
+| LLM&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; | Model&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; | Avg | SQA(I) | MMMU(Val) | MathVista | AI2D | Avg | MME | MMB | SEED(I) | GQA | Avg | ChartQA | OCR | TextVQA | DocVQA | Avg | MMVP | RWQA |
+| Llama 3-8B | Mini-Gemini-HD | 55.7 | 75.1 | 37.3 | 37.0 | 73.5 | 72.7 | 1606 | 72.7 | 73.2 | 64.5 | 62.9 | 59.1 | 47.7 | 70.2 | 74.6 | 40.4 | 18.7 | 62.1 |
+|            | LLaVA-NeXT     | 55.6 | 72.8 | 41.7 | 36.3 | 71.6 | 72.5 | 1604 | 72.1 | 72.7 | 65.2 | 63.9 | 69.5 | 49.0 | 64.6 | 72.6 | 49.4 | 38.7 | 60.1 |
+|            | Cambrian-1     | 61.3 | 80.4 | 42.7 | 49.0 | 73.0 | 73.1 | 1547 | 75.9 | 74.7 | 64.6 | 71.3 | 73.3 | 62.4 | 71.7 | 77.8 | 57.6 | 51.3 | 64.2 |
+|            | [Ealge-X4-8B-Plus](https://huggingface.co/NVEagle/Eagle-X4-8B-Plus) | 64.2 | 84.3 | 43.4 | 52.7 | 76.1 | 73.8 | 1559 | 75.9 | 76.3 | 64.9 | 76.6 | 80.1 | 62.6 | 77.1 | 86.6 | 69.1 | 71.6 | 66.5 |
+| Vicuna-13B | Mini-Gemini-HD | 54.1 | 71.9 | 37.3 | 37.0 | 70.1 | 70.7 | 1597 | 68.6 | 70.6 | 63.7 | 60.8 | 56.6 | 46.6 | 70.2 | 69.8 | 38.4 | 19.3 | 57.5 |
+|            | LLaVA-NeXT     | 53.7 | 73.5 | 36.2 | 35.1 | 70.0 | 69.9 | 1575 | 70.0 | 65.6 | 65.4 | 62.9 | 62.2 | 51.4 | 67.1 | 70.9 | 47.6 | 36.0 | 59.1 |
+|            | Cambrian-1     | 60.2 | 79.3 | 40.0 | 48.0 | 73.6 | 73.7 | 1610 | 75.7 | 74.4 | 64.3 | 71.3 | 73.8 | 61.9 | 72.8 | 76.8 | 52.2 | 41.3 | 63.0 |
+|            | [Ealge-X4-13B-Plus](https://huggingface.co/NVEagle/Eagle-X4-13B-Plus) | 63.0 | 82.0 | 41.0 | 54.4 | 74.0 | 74.6 | 1651 | 75.7 | 74.8 | 65.3 | 75.1 | 77.6 | 61.9 | 75.5 | 85.4 | 61.4 | 58.0 | 64.8 |
+|   Yi-34B   | Mini-Gemini-HD | 62.4 | 77.7 | 48.0 | 43.4 | 80.5 | 76.2 | 1659 | 80.6 | 75.3 | 65.8 | 68.1 | 67.6 | 51.8 | 74.1 | 78.9 | 52.3 | 37.3 | 67.2 |
+|            | LLaVA-NeXT     | 62.5 | 81.8 | 46.7 | 46.5 | 74.9 | 76.0 | 1633 | 79.3 | 75.9 | 67.1 | 67.7 | 68.7 | 54.5 | 69.5 | 78.1 | 54.2 | 47.3 | 61.0 |
+|            | Cambrian-1     | 67.0 | 85.6 | 49.7 | 53.2 | 79.7 | 76.8 | 1689 | 81.4 | 75.3 | 65.8 | 71.9 | 75.6 | 60.0 | 76.7 | 75.5 | 60.3 | 52.7 | 67.8 |
+|            | [Ealge-X5-34B-Plus](https://huggingface.co/NVEagle/Eagle-X5-34B-Plus) | 68.6 | 85.5 | 51.8 | 57.9 | 79.1 | 76.3 | 1677 | 81.0 | 75.6 | 64.9 | 75.4 | 77.2 | 62.4 | 78.8 | 83.0 | 68.3 | 67.0 | 69.5 |
+
+
+## Visual Examples
+
+### Knowledge & General VQA
+
+<div align="center">
+<img src="assets/visual/VQA1.png" width="80%">
+</div><br>
+
+<div align="center">
+<img src="assets/visual/VQA2.png" width="80%">
+</div><br>
+
+<div align="center">
+<img src="assets/visual/VQA3.png" width="80%">
+</div>
+
+### Autonomous Driving
+
+<div align="center">
+<img src="assets/visual/AV1.png" width="90%">
+</div><br>
+
+<div align="center">
+<img src="assets/visual/AV2.png" width="90%">
+</div>
+
+### Infographic, Chart, OCR & Document Understanding
+
+<div align="center">
+<img src="assets/visual/Doc1.png" width="80%">
+</div><br>
+
+<div align="center">
+<img src="assets/visual/Doc2.png" width="80%">
+</div><br>
+
+<div align="center">
+<img src="assets/visual/Doc3.png" width="80%">
+</div>
+
+
+## Install
+Please following the guide here to prepare the environment on **Linux OS**.
+<!-- currently does not support windows and MacOS -->
+
+1. Clone this repository
+```bash
+git clone https://github.com/NVlabs/EAGLE.git
+cd EAGLE
+```
+
+2. Create environment and install package
+```Shell
+conda create -n eagle python=3.10 -y
+conda activate eagle
+pip install --upgrade pip  # enable PEP 660 support
+pip install -r requirements.txt
+pip install .
+```
+
+3. Install additional packages for training cases
+```
+pip install flash-attn --no-build-isolation
+```
+If you have any questions about the environment setup, please follow the instruction [video](https://www.youtube.com/watch?si=20yjQlthlKPTC87s&v=0-md0S9GDJA&feature=youtu.be).
+
+
+## Training Data
+
+### Pre-training
+We use the same pretraining data as LLaVA v1.5, please download the data from [here](https://huggingface.co/datasets/liuhaotian/LLaVA-CC3M-Pretrain-595K).
+
+### Supervised Fine-tuning
+We have compiled all the data and images used in our supervised fine-tuning together. Please download the data from [here](https://huggingface.co/datasets/shi-labs/Eagle-1.8M).
+After cloning this dataset, please run the following commands to extract all the images:
+```
+cd Eagle-1.8M
+cat images.tar.part_* > images.tar.gz
+tar -xvf images.tar.gz
+```
+
+Please note that while the images have been packaged for convenience, the original dataset licenses remain unchanged. By downloading our data, you agree to the licensing terms of each source dataset. A detailed list of the data sources used in our fine-tuning data mixture is provided below:
+
+| Version         | Dataset Name    | Sample Number | Note                                 |
+|-----------------|-----------------|---------------|--------------------------------------|
+|                 | LLaVA v1.5      | 665k          | Multi-modal conversation             |
+|                 | DocVQA          | 39k           | Document understanding               |
+|                 | synDog-EN       | 50k           | OCR                                  |
+|                 | ChartQA         | 28k           | Chart understanding                  |
+|                 | DVQA            | 25k           | Chart understanding                  |
+|                 | AI2D            | 15k           | Diagram Understanding                |
+|                 | ShareGPT-4V     | 100k          | Detailed caption generated by GPT-4V |
+|                 | laion-gpt4v *   | 11k           | Detailed caption generated by GPT-4V |
+|                 | LVIS-Instruct4V | 220k          | Multi-modal conversation             |
+|                 | LRV-Instruct    | 150k          | Multi-modal conversation             |
+|                 | Geo170k         | 120k          | Math                                 |
+|                 | LLaVAR          | 20k           | OCR                                  |
+|                 | Visual7W        | 70k           | Visual Question Answering            |
+|                 | Open-Hermes 2.5 | 300k          | Text                                 |
+| Initial Version | Total           | 1.8M          |                                      |
+
+\* We have done manual inspection to ensure that the dataset does not contain any CSAM content.
+
+To pretrain or fine-tune our model on the [Cambrian-1](https://huggingface.co/collections/nyu-visionx/cambrian-data-6667ce801e179b4fbe774e11) dataset, please prepare the data according to their instructions. Then, convert the `jsonl` files into the `json` file by running the following python code:
+```python
+import json
+
+source_file = "Cambrian7M_withsystemprompt.jsonl"
+dst_file = "Cambrian7M_withsystemprompt.json"
+
+annotations = []
+with open(source_file, "r") as f:
+    for line in f:
+        annotations.append(json.loads(line))
+        
+with open(dst_file, "w") as f:
+    json.dump(annotations, f)
+```
+
+
+
+## Checkpoint Preparation
+Please provide the pretrained model weights for EVA-02 vision tower pretrained on detection task. You can download the checkpoint [here](https://huggingface.co/Yuxin-CV/EVA-02/blob/main/eva02/det/eva02_L_coco_det_sys_o365.pth) and place it in the `checkpoints/pretrained_models/` directory.
+
+The weights of other models, including Vicuna, Segment Anything Model, Pix2Struct, ConvNeXt, and CLIP will be automatically downloaded from huggingface during the first run.
+
+
+## Training
+
+The training process for Eagle follows a standard two-stage approach: pretraining and supervised fine-tuning. In the first stage, only the projector's weights are updated. In the second stage, all parameters are fine-tuned. The batch sizes for the pretraining and fine-tuning stages are 256 and 128, respectively. All settings and hyperparameters are identical to those in LLaVA-v1.5 except that we will unfrozen the vision tower's parameters during the second stage.
+
+In default we use 32 NVIDIA A100 80G GPU to conduct the training. Please modify the `per_device_train_batch_size` and `gradient_accumulation_steps` if you are using different amount of GPUs.
+
+### Pre-training
+If you are using a slurm cluster, please use the following command to submit a job.
+
+```
+srun \
+    --partition $your_partition \
+    --gres "gpu:8" \
+    --ntasks_per_node 1 \
+    -N 4 \
+    --job-name $RUN_NAME \
+    "bash $CMD $RUN_NAME"
+```
+
+You can specify the `RUN_NAME` and `CMD` variables to run different models according to the following table:
+| Model    | Language Model | Script                                    |
+|----------|----------------|-------------------------------------------|
+| Eagle-X4 | Vicuna-7B      | `scripts/pretrain-eagle-x4-vicuna-7b.sh`  |
+| Eagle-X4 | Vicuna-13B     | `scripts/pretrain-eagle-x4-vicuna-13b.sh` |
+| Eagle-X5 | Vicuna-7B      | `scripts/pretrain-eagle-x5-vicuna-7b.sh`  |
+| Eagle-X5 | Vicuna-13B     | `scripts/pretrain-eagle-x5-vicuna-13b.sh` |
+
+Remember to set the `$PATH_TO_PRETRAINING_DATA` in each script to the downloaded pretraining data. After you have complete the pretraining, you will get a file named `mm_projector.bin` in the checkpoint folder. 
+
+
+### Supervised Fine-tuning
+After pretraining is complete, a projector weight file `` will be saved in the checkpoint directory. Please set the `$PATH_TO_PRETRAINED_PROJECTOR` to the path of this projector weights.
+
+You can use the same sumbit code as the pretraining, and use the script in the following table to launch the supervised fine-tuning.
+| Model    | Language Model | Script                                         |
+|----------|----------------|------------------------------------------------|
+| Eagle-X4 | Vicuna-7B      | `scripts/finetune-eagle-x4-vicuna-7b-1.8m.sh`  |
+| Eagle-X4 | Vicuna-13B     | `scripts/finetune-eagle-x4-vicuna-13b-1.8m.sh` |
+| Eagle-X5 | Vicuna-7B      | `scripts/finetune-eagle-x5-vicuna-7b-1.8m.sh`  |
+| Eagle-X5 | Vicuna-13B     | `scripts/finetune-eagle-x5-vicuna-13b-1.8m.sh` |
+
+Before submit the job, you should correctly set the `$PATH_TO_SFT_DATA` and `$PATH_TO_PRETRAINED_PROJECTOR` in each script.
+
+### Notes
+If you have limited GPU resources or memory, please considering the following:
+
+- use `scripts/zero3.json` or `scripts/zero3_offload.json` as the Deepspeed training config instead of the default `zero2.json`
+- use gradient accumulation and reduce the per-device batch size
+
+
+## Inference
+Our inference code is [here](https://github.com/NVlabs/EAGLE/tree/main/predict_demo.py). You can set you own 'image_path' [here](https://github.com/NVlabs/EAGLE/tree/main/predict_demo.py/#L38) and 'question' [here](https://github.com/NVlabs/EAGLE/tree/main/predict_demo.py/#L39).
+```
+import os
+import torch
+import numpy as np
+from eagle import conversation as conversation_lib
+from eagle.constants import DEFAULT_IMAGE_TOKEN
+from eagle.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from eagle.conversation import conv_templates, SeparatorStyle
+from eagle.model.builder import load_pretrained_model
+from eagle.utils import disable_torch_init
+from eagle.mm_utils import tokenizer_image_token, get_model_name_from_path, process_images, KeywordsStoppingCriteria
+from PIL import Image
+import argparse
+from transformers import TextIteratorStreamer
+from threading import Thread
+
+model_path = "NVEagle/Eagle-X5-13B-Chat"
+conv_mode = "vicuna_v1"
+image_path = "assets/georgia-tech.jpeg"
+input_prompt = "Describe this image."
+
+model_name = get_model_name_from_path(model_path)
+tokenizer, model, image_processor, context_len = load_pretrained_model(model_path,None,model_name,False,False)
+if model.config.mm_use_im_start_end:
+    input_prompt = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + input_prompt
+else:
+    input_prompt = DEFAULT_IMAGE_TOKEN + '\n' + input_prompt
+
+conv = conv_templates[conv_mode].copy()
+conv.append_message(conv.roles[0], input_prompt)
+conv.append_message(conv.roles[1], None)
+prompt = conv.get_prompt()
+
+image = Image.open(image_path).convert('RGB')
+image_tensor = process_images([image], image_processor, model.config)[0]
+input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt')
+
+input_ids = input_ids.to(device='cuda', non_blocking=True)
+image_tensor = image_tensor.to(dtype=torch.float16, device='cuda', non_blocking=True)
+
+with torch.inference_mode():
+    output_ids = model.generate(
+        input_ids.unsqueeze(0),
+        images=image_tensor.unsqueeze(0),
+        image_sizes=[image.size],
+        do_sample=True,
+        temperature=0.2,
+        top_p=0.5,
+        num_beams=1,
+        max_new_tokens=256,
+        use_cache=True)
+
+outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
+print(f"Image:{image_path} \nPrompt:{input_prompt} \nOutput:{outputs}")
+```
+
+
+## Evaluation
+
+### Evaluation with LMMs-Eval
+We evaluate MME, MMBench, SEED, MathVista, POPE, ScienceQA, GQA, OCRBench, TextVQA, and ChartQA using [LMMs-Eval](https://github.com/EvolvingLMMs-Lab/lmms-eval). For better reproducibility, we have included the specific version we used in this repository. Please follow their guidelines and use the following commands to perform the evaluation:
+
+```bash
+bash scripts/eval_lmms_eval/eval-mme-seed-mmmu-pope-sqa-gqa-ocrbench-textvqa-chartqa.sh $REPO_ID_OR_LOCAL_PATH $MODEL_NAME $CONV_MODE
+# MODEL_NAME can be any name, just to dinstinguish different runs.
+# CONV_MODE should be the name of the conversation template during triaining, i.e., "vicuna_v1" for Vicuna, "llama3" for Llama3, and "yi_34b_chatml_direct" for Yi-34B.
+```
+
+
+## Gradio Demo
+We set up an online demo [here](https://huggingface.co/spaces/NVEagle/Eagle-X5-13B-Chat). You can also run this demo on your own machine by running:
+```
+python gradio_demo.py \
+    --model-path ${MODEL_CKPT}
+    --conv-mode vicuna_v1
+```
+
+
+## Citation
+If you find this project useful, please cite our work:
+```
+@article{shi2024eagle,
+    title = {Eagle: Exploring The Design Space for Multimodal LLMs with Mixture of Encoders}, 
+    author={Min Shi and Fuxiao Liu and Shihao Wang and Shijia Liao and Subhashree Radhakrishnan and De-An Huang and Hongxu Yin and Karan Sapra and Yaser Yacoob and Humphrey Shi and Bryan Catanzaro and Andrew Tao and Jan Kautz and Zhiding Yu and Guilin Liu},
+    journal={arXiv:2408.15998},
+    year={2024}
+}
+```
+
+
+## License
+[![Code License](https://img.shields.io/badge/Code%20License-Apache_2.0-green.svg)](https://github.com/tatsu-lab/stanford_alpaca/blob/main/LICENSE)
+**Usage and License Notices**: This project utilizes certain datasets and checkpoints that are subject to their respective original licenses. Users must comply with all terms and conditions of these original licenses, including but not limited to the [OpenAI Terms of Use](https://openai.com/policies/terms-of-use) for the dataset and the specific licenses for base language models for checkpoints trained using the dataset (e.g. [Llama community license](https://ai.meta.com/llama/license/) for Llama-2, Llama-3, and Vicuna-v1.5). This project does not impose any additional constraints beyond those stipulated in the original licenses. Furthermore, users are reminded to ensure that their use of the dataset and checkpoints is in compliance with all applicable laws and regulations.
+
+
+## Acknowledgement
+
+- [LLaVA](https://github.com/haotian-liu/LLaVA): the codebase we built upon. Thanks for the great pioneer open-source project!
+- [LMMs-Eval](https://github.com/EvolvingLMMs-Lab/lmms-eval): many thanks to the LMMs-Lab for their wonderful and easy-to-use evaluation tools!
+- [LLaVA-HR](https://github.com/luogen1996/LLaVA-HR): we borrow some code on flexible input CLIP encoder from LLaVA-HR!
+- [Cambrian-1](https://cambrian-mllm.github.io): thanks Cambrian project contributors for their efforts in organizing open-source data for us!
+- Thanks to the [VILA](https://github.com/NVlabs/VILA) team and the [RADIO](https://github.com/NVlabs/RADIO) team for their helps and discussions. Check out these awesome works from NVIDIA!

EAGLE/evaluate_lmms_eval.py

@@ -0,0 +1,345 @@
+import os
+import yaml
+import sys
+import copy
+import json
+import logging
+import traceback
+import argparse
+import torch
+import numpy as np
+import datetime
+
+import warnings
+import traceback
+
+warnings.simplefilter("ignore", category=DeprecationWarning)
+
+from accelerate import Accelerator
+from accelerate.utils import InitProcessGroupKwargs
+from pathlib import Path
+from typing import Union
+import hashlib
+
+from lmms_eval import evaluator, utils
+from lmms_eval.tasks import initialize_tasks, include_path, get_task_dict
+from lmms_eval.api.registry import ALL_TASKS
+from lmms_eval.logging_utils import WandbLogger
+from lmms_eval.utils import PathFormatter
+
+
+eval_logger = logging.getLogger("lmms-eval")
+
+
+def _handle_non_serializable(o):
+    if isinstance(o, np.int64) or isinstance(o, np.int32):
+        return int(o)
+    elif isinstance(o, set):
+        return list(o)
+    else:
+        return str(o)
+
+
+def parse_eval_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(formatter_class=argparse.RawTextHelpFormatter)
+    parser.add_argument("--config", default="", help="Path to a yaml file specifying all eval arguments, will ignore cli arguments if specified")
+    parser.add_argument("--model", default="hf", help="Name of model e.g. `hf`")
+    parser.add_argument(
+        "--tasks",
+        default=None,
+        help="To get full list of tasks, use the command lmms-eval --tasks list",
+    )
+    parser.add_argument(
+        "--model_args",
+        default="",
+        help="String arguments for model, e.g. `pretrained=EleutherAI/pythia-160m,dtype=float32`",
+    )
+    parser.add_argument(
+        "--num_fewshot",
+        type=int,
+        default=None,
+        help="Number of examples in few-shot context",
+    )
+    parser.add_argument("--batch_size", type=str, default=1)
+    parser.add_argument(
+        "--device",
+        type=str,
+        default=None,
+        help="Device to use (e.g. cuda, cuda:0, cpu)",
+    )
+    parser.add_argument(
+        "--output_path",
+        default=None,
+        type=str,
+        metavar="= [dir/file.jsonl] [DIR]",
+        help="The path to the output file where the result metrics will be saved. If the path is a directory and log_samples is true, the results will be saved in the directory. Else the parent directory will be used.",
+    )
+    parser.add_argument(
+        "--limit",
+        type=float,
+        default=None,
+        help="Limit the number of examples per task. " "If <1, limit is a percentage of the total number of examples.",
+    )
+    parser.add_argument(
+        "--check_integrity",
+        action="store_true",
+        help="Whether to run the relevant part of the test suite for the tasks",
+    )
+    parser.add_argument(
+        "--show_task_to_terminal",
+        action="store_true",
+        default=False,
+        help="Prints the prompt for the first few documents",
+    )
+    parser.add_argument(
+        "--log_samples",
+        action="store_true",
+        default=False,
+        help="If True, write out all model outputs and documents for per-sample measurement and post-hoc analysis",
+    )
+    parser.add_argument(
+        "--wandb_log_samples",
+        action="store_true",
+        default=False,
+        help="If True, write out all model outputs and documents for per-sample measurement and post-hoc analysis to Weights and Biases",
+    )
+    parser.add_argument(
+        "--log_samples_suffix",
+        type=str,
+        default="",
+        help="Specify a suffix for the log_samples file name.",
+    )
+    parser.add_argument(
+        "--show_config",
+        action="store_true",
+        default=False,
+        help="If True, shows the the full config of all tasks at the end of the evaluation.",
+    )
+    parser.add_argument(
+        "--include_path",
+        type=str,
+        default=None,
+        help="Additional path to include if there are external tasks to include.",
+    )
+    parser.add_argument(
+        "--gen_kwargs",
+        default="",
+        help=("String arguments for model generation on greedy_until tasks," " e.g. `temperature=0,top_k=0,top_p=0`"),
+    )
+    parser.add_argument(
+        "--verbosity",
+        type=str,
+        default="INFO",
+        help="Log error when tasks are not registered.",
+    )
+    parser.add_argument(
+        "--wandb_args",
+        default="",
+        help="Comma separated string arguments passed to wandb.init, e.g. `project=lmms-eval,job_type=eval",
+    )
+    parser.add_argument(
+        "--timezone",
+        default="Asia/Singapore",
+        help="Timezone for datetime string, e.g. Asia/Singapore, America/New_York, America/Los_Angeles",
+    )
+    args = parser.parse_args()
+    return args
+
+
+def cli_evaluate(args: Union[argparse.Namespace, None] = None) -> None:
+    if not args:
+        args = parse_eval_args()
+
+    # Check if no arguments were passed after parsing
+    if len(sys.argv) == 1:
+        print("┌───────────────────────────────────────────────────────────────────────────────┐")
+        print("│ Please provide arguments to evaluate the model. e.g.                          │")
+        print("│ `lmms-eval --model llava --model_path liuhaotian/llava-v1.6-7b --tasks okvqa` │")
+        print("│ Use `lmms-eval --help` for more information.                                  │")
+        print("└───────────────────────────────────────────────────────────────────────────────┘")
+        sys.exit(1)
+
+    set_loggers(args)
+    eval_logger = logging.getLogger("lmms-eval")
+    eval_logger.setLevel(getattr(logging, f"{args.verbosity}"))
+    eval_logger.info(f"Verbosity set to {args.verbosity}")
+    os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+    args_list = []
+    results_list = []
+    if args.config:
+        if not os.path.exists(args.config):
+            raise ValueError(f"Config file does not exist: {args.config}")
+
+        with open(args.config, "r") as file:
+            config_args = yaml.safe_load(file)
+        config_args = [config_args] if type(config_args) != list else config_args
+        # multiple configs, create args list first
+        for config in config_args:
+            args_copy = argparse.Namespace(**vars(args))
+            for key, value in config.items():
+                setattr(args_copy, key, value)
+            args_list.append(args_copy)
+    else:
+        args_list.append(args)
+
+    # initialize Accelerator
+    kwargs_handler = InitProcessGroupKwargs(timeout=datetime.timedelta(seconds=60000))
+    accelerator = Accelerator(kwargs_handlers=[kwargs_handler])
+    if accelerator.is_main_process:
+        is_main_process = True
+    else:
+        is_main_process = False
+
+    for args in args_list:
+        try:
+            if is_main_process and args.wandb_args:  # thoughtfully we should only init wandb once, instead of multiple ranks to avoid network traffics and unwanted behaviors.
+                wandb_logger = WandbLogger(args)
+
+            # TO REMOVE
+            print(args)
+            results, samples = cli_evaluate_single(args)
+            results_list.append(results)
+
+            accelerator.wait_for_everyone()
+            if is_main_process and args.wandb_args:
+                wandb_logger.post_init(results)
+                wandb_logger.log_eval_result()
+                if args.wandb_log_samples and samples is not None:
+                    wandb_logger.log_eval_samples(samples)
+
+                wandb_logger.finish()
+
+        except Exception as e:
+            traceback.print_exc()
+            eval_logger.error(f"Error during evaluation: {e}")
+            traceback.print_exc()
+            results_list.append(None)
+
+    for args, results in zip(args_list, results_list):
+        # cli_evaluate will return none if the process is not the main process (rank 0)
+        if results is not None:
+            print_results(args, results)
+
+
+def cli_evaluate_single(args: Union[argparse.Namespace, None] = None) -> None:
+    eval_logger = logging.getLogger("lmms-eval")
+    eval_logger.setLevel(getattr(logging, f"{args.verbosity}"))
+    eval_logger.info(f"Verbosity set to {args.verbosity}")
+    os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+    initialize_tasks(args.verbosity)
+
+    if args.limit:
+        eval_logger.warning(" --limit SHOULD ONLY BE USED FOR TESTING." "REAL METRICS SHOULD NOT BE COMPUTED USING LIMIT.")
+    if args.include_path is not None:
+        eval_logger.info(f"Including path: {args.include_path}")
+        include_path(args.include_path)
+
+    if args.tasks is None:
+        task_names = ALL_TASKS
+    elif args.tasks == "list":
+        eval_logger.info("Available Tasks:\n - {}".format(f"\n - ".join(sorted(ALL_TASKS))))
+        sys.exit()
+    elif args.tasks == "list_with_num":
+        log_message = (
+            "\n" + "=" * 70 + "\n" + "\n\tYou are trying to check all the numbers in each task." + "\n\tThis action will download the complete dataset." + "\n\tIf the results are not clear initially, call this again." + "\n\n" + "=" * 70
+        )
+        eval_logger.info(log_message)
+        task_dict = get_task_dict([task for task in sorted(ALL_TASKS)], model_name="llava")
+        for task_name in task_dict.keys():
+            task_obj = task_dict[task_name]
+            if type(task_obj) == tuple:
+                group, task_obj = task_obj
+                if task_obj is None:
+                    continue
+            eval_logger.info(f"\nTask : {task_obj.config.task}\n - #num : {len(task_obj.test_docs()) if task_obj.has_test_docs() else len(task_obj.validation_docs())}")
+        sys.exit()
+    else:
+        tasks_list = args.tasks.split(",")
+        eval_logger.info(f"Evaluating on {len(tasks_list)} tasks.")
+        task_names = utils.pattern_match(tasks_list, ALL_TASKS)
+        task_missing = [task for task in tasks_list if task not in task_names and "*" not in task]  # we don't want errors if a wildcard ("*") task name was used
+
+        if task_missing:
+            missing = ", ".join(task_missing)
+            eval_logger.error(
+                f"Tasks were not found: {missing}. Try `lmms-eval --tasks list` for list of available tasks",
+            )
+            # eval_logger.warn(f"Tasks {missing} were not found. Try `lmms-eval --tasks list` for list of available tasks.")
+
+    eval_logger.info(f"Selected Tasks: {task_names}")
+
+    # set datetime before evaluation
+    datetime_str = utils.get_datetime_str(timezone=args.timezone)
+    if args.output_path:
+        hash_input = f"{args.model_args}".encode("utf-8")
+        hash_output = hashlib.sha256(hash_input).hexdigest()[:6]
+        path = Path(args.output_path)
+        path = path.expanduser().resolve().joinpath(f"{datetime_str}_{args.log_samples_suffix}_{args.model}_model_args_{hash_output}")
+        args.output_path = path
+
+    elif args.log_samples and not args.output_path:
+        assert args.output_path, "Specify --output_path"
+
+    results = evaluator.simple_evaluate(
+        model=args.model,
+        model_args=args.model_args,
+        tasks=task_names,
+        num_fewshot=args.num_fewshot,
+        batch_size=args.batch_size,
+        device=args.device,
+        limit=args.limit,
+        check_integrity=args.check_integrity,
+        show_task_to_terminal=args.show_task_to_terminal,
+        log_samples=args.log_samples,
+        gen_kwargs=args.gen_kwargs,
+        cli_args=args,
+    )
+
+    if results is not None:
+        if args.log_samples:
+            samples = results.pop("samples")
+        else:
+            samples = None
+        dumped = json.dumps(results, indent=4, default=_handle_non_serializable)
+        if args.show_config:
+            print(dumped)
+
+        if args.output_path:
+            args.output_path.mkdir(parents=True, exist_ok=True)
+            result_file_path = path.joinpath("results.json")
+            if result_file_path.exists():
+                eval_logger.warning(f"Output file {result_file_path} already exists and will be overwritten.")
+
+            result_file_path.open("w").write(dumped)
+            if args.log_samples:
+                for task_name, config in results["configs"].items():
+                    filename = args.output_path.joinpath(f"{task_name}.json")
+                    # Structure the data with 'args' and 'logs' keys
+                    data_to_dump = {"args": vars(args), "model_configs": config, "logs": sorted(samples[task_name], key=lambda x: x["doc_id"])}  # Convert Namespace to dict
+                    samples_dumped = json.dumps(data_to_dump, indent=4, default=_handle_non_serializable)
+                    filename.open("w").write(samples_dumped)
+                    eval_logger.info(f"Saved samples to {filename}")
+
+        return results, samples
+    return None, None
+
+
+def print_results(args, results):
+    print(f"{args.model} ({args.model_args}),\ngen_kwargs: ({args.gen_kwargs}),\nlimit: {args.limit},\nnum_fewshot: {args.num_fewshot},\nbatch_size: {args.batch_size}")
+    print(evaluator.make_table(results))
+    if "groups" in results:
+        print(evaluator.make_table(results, "groups"))
+
+
+def set_loggers(args):
+    eval_logger = logging.getLogger("lmms-eval")
+    ch = logging.StreamHandler()
+    formatter = PathFormatter("%(asctime)s [%(pathname)s:%(lineno)d] %(levelname)s %(message)s", "%m-%d %H:%M:%S", timezone=args.timezone)
+    ch.setFormatter(formatter)
+    eval_logger.addHandler(ch)
+
+
+if __name__ == "__main__":
+    cli_evaluate()

Emu/README.md

@@ -0,0 +1,55 @@
+<div align="center">
+
+<h2>Emu: Generative Multimodal Models from BAAI</h2>
+
+
+</div>
+
+---
+
+<div align='center'>
+<img src="./Emu2/assets/short_teaser.png" class="interpolation-image" alt="comparison_fewshot." height="80%" width="80%" />
+</div>
+
+
+
+- [**Emu1**](Emu1) (ICLR 2024, 2023/07) - Generative Pretraining in Multimodality
+
+- [**Emu2**](Emu2) (CVPR 2024, 2023/12) - Generative Multimodal Models are In-Context Learners
+
+- [**Emu3**](https://github.com/baaivision/Emu3) (arXiv 2024, 2024/09) - Next-Token Prediction is All You Need 🔥🔥🔥
+
+## News
+- 2024.9 We introduce **Emu3**, a new suite of state-of-the-art multimodal models trained solely with next-token prediction. 🔥🔥🔥
+- 2024.2 **Emu1 and Emu2 are accepted by ICLR 2024 and CVPR 2024 respectively! 🎉**
+- 2023.12 Inference code, model and demo of Emu2 are available. Enjoy the [demo](http://218.91.113.230:9002/).
+- 2023.12 We have released Emu2, open and largest generative multimodal models that achieve new state of the art on multimodal understanding and generation tasks.
+- 2023.7 Inference code and model of Emu are available.
+- 2023.7 We have released Emu, a multimodal generalist that can seamlessly generate images and texts in multimodal context.
+
+
+## Hightlights
+- State-of-the-art performance
+- Next-generation capabilities
+- A base model for diverse tasks
+
+We hope to foster the growth of our community through open-sourcing and promoting collaboration👬. Let's step towards multimodal intelligence together🍻.
+
+
+## Contact
+- **We are hiring** at all levels at BAAI Vision Team, including full-time researchers, engineers and interns. 
+If you are interested in working with us on **foundation model, visual perception and multimodal learning**, please contact [Xinlong Wang](https://www.xloong.wang/) (`[email protected]`).
+
+
+## Misc
+
+<div align="center">
+
+[![Stargazers repo roster for @baaivision/Emu](https://bytecrank.com/nastyox/reporoster/php/stargazersSVG.php?user=baaivision&repo=Emu)](https://github.com/baaivision/Emu/stargazers)
+
+
+[![Forkers repo roster for @baaivision/Emu](https://bytecrank.com/nastyox/reporoster/php/forkersSVG.php?user=baaivision&repo=Emu)](https://github.com/baaivision/Emu/network/members)
+
+[![Star History Chart](https://api.star-history.com/svg?repos=baaivision/Emu&type=Date)](https://star-history.com/#baaivision/Emu&Date)
+
+</div>

LLM2CLIP/CODE_OF_CONDUCT.md

@@ -0,0 +1,9 @@
+# Microsoft Open Source Code of Conduct
+
+This project has adopted the [Microsoft Open Source Code of Conduct](https://opensource.microsoft.com/codeofconduct/).
+
+Resources:
+
+- [Microsoft Open Source Code of Conduct](https://opensource.microsoft.com/codeofconduct/)
+- [Microsoft Code of Conduct FAQ](https://opensource.microsoft.com/codeofconduct/faq/)
+- Contact [[email protected]](mailto:[email protected]) with questions or concerns

LLaVA/.gitattributes

@@ -0,0 +1,29 @@
+# https://git-scm.com/docs/gitattributes
+
+# Set the default behavior, in case people don't have core.autocrlf set.
+# https://git-scm.com/docs/gitattributes#_end_of_line_conversion
+* text=auto
+
+# common python attributes, taken from https://github.com/alexkaratarakis/gitattributes/blob/710900479a2bedeec7003d381719521ffbb18bf8/Python.gitattributes
+# Source files
+# ============
+*.pxd    text diff=python
+*.py     text diff=python
+*.py3    text diff=python
+*.pyw    text diff=python
+*.pyx    text diff=python
+*.pyz    text diff=python
+*.pyi    text diff=python
+
+# Binary files
+# ============
+*.db     binary
+*.p      binary
+*.pkl    binary
+*.pickle binary
+*.pyc    binary export-ignore
+*.pyo    binary export-ignore
+*.pyd    binary
+
+# Jupyter notebook
+*.ipynb  text eol=lf

LLaVA/.gitignore

@@ -0,0 +1,35 @@
+# Python
+__pycache__
+*.pyc
+*.egg-info
+dist
+
+# Log
+*.log
+*.log.*
+*.json
+*.jsonl
+
+# Data
+!**/alpaca-data-conversation.json
+
+# Editor
+.idea
+*.swp
+
+# Other
+.DS_Store
+wandb
+output
+
+checkpoints
+ckpts*
+
+.ipynb_checkpoints
+*.ipynb
+
+# DevContainer
+!.devcontainer/*
+
+# Demo
+serve_images/

PaddleMIX/.gitmodules

@@ -0,0 +1,3 @@
+[submodule "PaddleNLP"]
+	path = PaddleNLP
+	url = https://github.com/PaddlePaddle/PaddleNLP

PaddleMIX/CITATION.cff

@@ -0,0 +1,15 @@
+# This CITATION.cff file was generated with cffinit.
+# Visit https://bit.ly/cffinit to generate yours today!
+
+cff-version: 1.2.0
+title: Paddle Multimodal Integration and eXploration
+message: >-
+  If you use this repository, please cite it using the metadata from this file.
+type: software
+authors:
+  - given-names: PaddleMIX Authors
+repository-code: 'https://github.com/PaddlePaddle/PaddleMIX'
+repository: 'https://github.com/PaddlePaddle/PaddleMIX'
+keywords:
+  - paddlemix
+license: Apache-2.0

PaddleMIX/README.md

@@ -0,0 +1,413 @@
+简体中文 | [English](README_EN.md)
+
+<p align="center">
+  <img src="https://github.com/PaddlePaddle/PaddleMIX/assets/22989727/2cd19298-1c52-4d73-a0f7-dcdab6a8ec90" align="middle" width = "600" />
+</p>
+
+<p align="center">
+    <a href="https://github.com/PaddlePaddle/PaddleMix/releases"><img src="https://img.shields.io/github/v/release/PaddlePaddle/PaddleMix?color=ffa"></a>
+    <a href="./LICENSE"><img src="https://img.shields.io/badge/license-Apache%202-dfd.svg"></a>
+    <a href=""><img src="https://img.shields.io/badge/python-3.7+-aff.svg"></a>
+    <a href=""><img src="https://img.shields.io/badge/os-linux%2C%20win%2C%20mac-pink.svg"></a>
+    <a href="#📌社区交流"><img src="https://img.shields.io/badge/微信-小助手加群-green?logo=wechat&amp"></a>
+    <a href="https://github.com/PaddlePaddle/PaddleMIX/stargazers"><img src="https://img.shields.io/github/stars/PaddlePaddle/PaddleMIX?color=ccf"></a>
+
+</p>
+</div>
+
+## 💌目录
+- [💌目录](#目录)
+- [📰新闻](#新闻)
+- [📣最新进展](#最新进展)
+- [🌈简介](#简介)
+- [✨主要特性](#主要特性)
+    - [📱丰富的多模态功能](#丰富的多模态功能)
+    - [🧩简洁的开发体验](#简洁的开发体验)
+    - [💡高性能分布式训推能力](#高性能分布式训推能力)
+    - [🔧特色功能与工具](#特色功能与工具)
+- [🔍安装](#安装)
+- [🔥教程](#教程)
+- [🤔FAQ](#faq)
+- [📱模型库](#模型库)
+- [📝许可证书](#许可证书)
+- [📌社区交流](#社区交流)
+
+
+## 📰新闻
+**🔥2024.11.21日 - 2024.12.22日  PaddleMIX开发项目挑战（已结束）**
+
+- ✨「体验官招募」PaddleMIX开发项目挑战
+点击链接报名🔗：https://aistudio.baidu.com/activitydetail/1503019366
+🏆投稿至飞桨星河社区项目大厅，加精获得PaddleMIX体验官认证证书及京东卡激励
+欢迎大家投稿～
+<details>
+<summary>点击展开活动海报</summary>
+<p align="center">
+<img src='https://github.com/user-attachments/assets/27e0bbe3-0ff8-49ef-bd39-81a31a2b288b'  width="25%">
+</p>
+</details>
+
+## 📣最新进展
+
+<!-- 📚《飞桨多模态大模型开发套件PaddleMIX 2.1 震撼发布》，图文音视频场景全覆盖，多模态高效助力产业创新。超大规模训练支持，覆盖图文预训练、文生图、跨模态视觉任务，覆盖金融、教育、电商、医疗等产业场景。8月8日（周四）20：00 带你直播了解多模态大模型最新架构，深度解析PaddleMIX高性能模型库，手把手演示LLaVA模型训推全流程。[报名链接](https://www.wjx.top/vm/wKqysjx.aspx?udsid=449688)   -->
+
+
+**🎉 2024.12.17 支持[GOT-OCR2_0](./paddlemix/examples/GOT_OCR_2_0)推理和训练**
+
+**🎉 2024.12.17 支持[InternVL2_5(1B、2B、4B、8B)](./paddlemix/examples/internvl2)推理**
+
+**🎉 2024.11.27 支持[Janus/JanusFlow](./paddlemix/examples/janus)推理**
+
+**🎉 2024.11.21 支持[MiniCPM-V-2_6](./paddlemix/examples/minicpm-v-2_6)推理**
+
+**🎉 2024.11.8 支持[DenseConnector](./paddlemix/examples/llava_denseconnector)和[Aquila-VL-2B-llava-qwen](./paddlemix/examples/llava_onevision/)推理**
+
+**🎉 2024.11.1 支持[LLaVA-OneVision](./paddlemix/examples/llava_onevision/)和[LLaVA-Critic](./paddlemix/examples/llava_critic/)推理**
+
+**🎉 2024.10.31 喜迎外部开发者的[创作教程页面](paddlemix_applications.md)更新**
+
+* 🌟 自9月6日发起大模型套件精品项目征集活动以来,我们收到了30个优质开发者项目,其中25个精品项目已通过平台评估并成功加精。
+
+* 🙏 衷心感谢各位开发者基于套件的精彩创作！🚀 诚挚邀请您也来分享您的创意 - 欢迎将教程发布到公开网页或[飞桨AI Studio](https://aistudio.baidu.com/aistudio/community/multimodal?from=singlemessage)社区！
+
+<details>
+<summary>点击展开更多</summary>
+
+**🔥2024.10.11 发布PaddleMIX v2.1**
+* 支持[PaddleNLP 3.0 beta](https://github.com/PaddlePaddle/PaddleNLP/releases/tag/v3.0.0-beta0)版本，抢先体验其最新功能。
+* 新增[Qwen2-VL](./paddlemix/examples/qwen2_vl/)、[InternVL2](./paddlemix/examples/internvl2/)、[Stable Diffusion 3 (SD3)](https://github.com/PaddlePaddle/PaddleMIX/blob/develop/ppdiffusers/examples/dreambooth/README_sd3.md)等前沿模型。
+* 发布自研多模数据能力标签模型[PP-InsCapTagger](./paddlemix/datacopilot/example/pp_inscaptagger/)；可用于数据的分析和过滤，试验案例表明在保持模型效果的条件下可减少50%的数据量，大幅提高训练效率。
+
+* 多模态大模型InternVL2、LLaVA、SD3、SDXL适配昇腾910B，提供国产计算芯片上的训推能力。
+
+
+**2024.07.25 发布PaddleMIX v2.0**
+* 多模态理解：新增LLaVA系列，Qwen-VL等；新增Auto模块统一SFT训练流程；新增mixtoken训练策略，SFT吞吐量提升5.6倍。
+* 多模态生成：发布[PPDiffusers 0.24.1](./ppdiffusers/README.md)版本，支持视频生成能力，文生图模型新增LCM。新增飞桨版peft，accelerate后端。提供基于飞桨开发的ComfyUI插件。
+* 多模态数据处理工具箱[DataCopilot](./paddlemix/datacopilot/)：支持自定义数据结构，数据转换，离线格式检查；支持基本的统计信息，数据可视化功能。
+
+**2023.10.7 发布 PaddleMIX v1.0**
+* 新增图文预训练模型分布式训练能力，BLIP-2支持千亿规模训练
+* 新增跨模态应用流水线[AppFlow](./applications/README.md)，一键支持自动标注，图像编辑，音生图等11种跨模态应用
+* [PPDiffusers](./ppdiffusers/README.md)发布 0.19.3 版本，新增SDXL及相关任务
+
+</details>
+
+---
+
+## 🌈简介
+
+PaddleMIX是基于飞桨的多模态大模型开发套件，聚合图像、文本、视频等多种模态，覆盖视觉语言预训练，微调，文生图，文生视频，多模态理解等丰富的多模态任务。它提供开箱即用的开发体验，同时支持灵活定制，满足不同需求，助力探索通用人工智能。
+
+<p align="center">
+  <img src="https://github.com/user-attachments/assets/764b32a4-3933-4ef8-a0b2-dd425af49ef8" align="middle" width = 100% />
+</p>
+
+PaddleMIX工具链包括数据处理、模型开发、预训练、精调和推理部署，支持主流多模态模型如 EVA-CLIP、BLIP-2、Stable Diffusion 等。通过跨模态任务流水线 AppFlow 和文生图应用 pipeline，开发者可以快速构建多模态应用。
+
+### 多模态理解效果示例如下：
+
+<img src="https://github.com/user-attachments/assets/4c9a0427-57c7-4e1b-80f0-428c03119cc3"></img>
+
+
+多模态理解🤝融合了视觉👀和语言💬处理能力。包含基础感知、细粒度图像理解和复杂视觉推理🧠等功能。我们的[模型库](#模型库)调用提供了单图、多图和视频推理的功能实际应用，功能包括自然图像摘要📝、问答🤔、OCR🔍、情感识别❤️😢、专业图像分析🔬和代码解析💻。这些技术可应用于教育📚、医疗🏥、工业🏭等多个领域，实现从静态图像🖼️到动态视频🎥的全面智能分析。欢迎您的体验和探索～
+
+### 多模态生成效果示例如下：
+<div style="display: flex; justify-content: center; gap: 5px;">
+    <img src="https://github.com/user-attachments/assets/f4768f08-f7a3-45e0-802c-c91554dc5dfc" style="height: 250px; object-fit: fill;">
+    <img src="https://github.com/user-attachments/assets/9bf4a333-af57-4ddd-a514-617dea8da435" style="height: 250px; object-fit: fill;">
+</div>
+
+
+多模态生成✍️融合了文本💬与视觉👀的创造能力。涵盖了从文字生成图像🖼️到文字生成视频🎥的各类技术，包括 Stable Diffusion 3、Open-Sora等先进模型。我们在[ppdiffusers](ppdiffusers/README.md)提供了单图生成、多图合成和视频生成的实际应用，功能涉及艺术创作🎨、动画制作📽️、内容生成📝等。通过这些技术，可以在教育📚、娱乐🎮、广告📺等领域实现从静态图像到动态视频的创意生成。欢迎您的体验和探索～
+
+### 特色应用效果示例如下（点击标题可快速跳转在线体验）：
+|                                                  [**ComfyUI创作工作流**](https://aistudio.baidu.com/community/app/106043)                                                  |                                                [**艺术风格二维码模型**](https://aistudio.baidu.com/community/app/1339)                                                |                                                  [**Mix叠图**](https://aistudio.baidu.com/community/app/1340)                                                  |
+| :--------------------------------------------------------------------------------------------------------------------------------------------: | :----------------------------------------------------------------------------------------------------------------------------------------------: | :--------------------------------------------------------------------------------------------------------------------------------------: |
+| <img src='https://github.com/PaddlePaddle/PaddleMIX/assets/35400185/36ba7261-1744-41a4-b1cb-c9e99f6931f2' width="300px"> | <img src='https://github.com/PaddlePaddle/Paddle/assets/22989727/ba091291-a1ee-49dc-a1af-fc501c62bfc8'  width="300px"> | <img src='https://github.com/PaddlePaddle/Paddle/assets/22989727/a71be5a0-b0f3-4aa8-bc20-740ea8ae6785'  width="300px"> |
+|                                                  [**二次元文生图**](https://aistudio.baidu.com/community/app/2/webUI?source=appCenter)                                                   |                                                     [**AI绘画｜50+Lora风格叠加**](https://aistudio.baidu.com/community/app/2848/webUI?source=appCenter)                                                     |                                               [**ControlNet｜图片局部重绘**](https://aistudio.baidu.com/community/app/1981/webUI?source=appCenter)                                               |
+| <img src='https://github.com/user-attachments/assets/a4af8f8a-08c7-4da7-8575-9dbfedaba56c' width="200px"> | <img src='https://github.com/user-attachments/assets/fa92c229-a885-46a1-b23f-a076855c93ec'  width="200px"> | <img src='https://github.com/user-attachments/assets/78625876-d8ec-4c15-ae96-655c50f562ab'  width="200px"> |
+
+
+
+
+
+-----
+
+
+
+
+
+
+
+
+## ✨主要特性
+
+### 📱丰富的多模态功能
+PaddleMIX支持大量最新主流的算法基准以及预训练模型，覆盖图文预训练，文生图，跨模态视觉任务，实现图像编辑、图像描述、数据标注等多样功能。`传送门`：[📱模型库](#模型库)
+
+### 🧩简洁的开发体验
+PaddleMIX 提供统一的模型开发接口，支持开发者快速集成和定制模型。借助 Auto 模块，用户可以高效加载预训练模型、实现 Tokenization，并通过简化的 API 轻松完成模型的训练、微调（SFT）、推理与部署。此外，Auto 模块支持开发者自定义模型的自动化集成，确保灵活性与可扩展性，同时提升开发效率。
+
+### 💡高性能分布式训推能力
+PaddleMIX提供高性能分布式训练与推理能力，融合✨Fused Linear✨、✨Flash Attention✨等加速算子，支持🌀BF16混合精度训练和4D混合并行策略，并通过优化推理性能，包括卷积布局、GroupNorm融合及旋转位置编码优化，显著提升大规模预训练和高效推理性能。
+
+<img src="https://github.com/user-attachments/assets/9ab9540a-fa89-41cb-838d-95df86e33382" width = 100% />
+
+
+
+### 🔧特色功能与工具
+多模态数据处理工具箱DataCopilot，加速模型迭代升级。让开发者根据特定任务以低代码量实现数据的基本操作。`传送门`：[🏆特色模型|工具](#特色模型工具)
+
+
+## 🔍安装
+### 1. 克隆PaddleMIX仓库
+```
+git clone https://github.com/PaddlePaddle/PaddleMIX
+cd PaddleMIX
+```
+
+### 2. 创建虚拟环境
+```
+conda create -n paddlemix python=3.10 -y
+conda activate paddlemix
+```
+### 3. ‼️安装PaddlePaddle
+
+#### 方法 1: 一键安装（GPU/CPU推荐）
+
+- CUDA 11.x或12.3
+- PaddlePaddle 3.0.0b1
+```
+sh build_paddle_env.sh
+```
+
+#### 方法 2: 手动安装
+关于PaddlePaddle安装的详细教程请查看[Installation](https://www.paddlepaddle.org.cn/install/quick?docurl=/documentation/docs/zh/develop/install/pip/linux-pip.html)。
+
+### 4. ‼️安装依赖
+
+#### 方法 1: 一键安装（推荐）
+
+运行以下命令来自动安装所有必要的依赖:
+```
+sh build_env.sh
+```
+
+#### 方法 2: 手动安装
+```bash
+# 安装 PaddleMIX
+pip install -e .
+# 安装 ppdiffusers
+cd ppdiffusers
+pip install -e .
+cd ..
+```
+### 5. ‼️验证安装
+```bash
+sh check_env.sh
+
+环境和依赖推荐版本:
+- paddlepaddle: 3.0.0b2或develop版本
+- paddlenlp: 3.0.0b2
+- ppdiffusers: 0.29.0
+- huggingface_hub: 0.23.0
+```
+
+### 6. 安装自定义算子（可选）
+* 部分模型需要安装自定义算子（FastLayerNorm、FusedLayerNorm），例如EVA-CLIP、DIT_LLAMA等。
+* 非CUDA环境（例如昇腾环境）则跳过
+```bash
+cd paddlemix/external_ops
+python setup.py install
+```
+
+## 🔥教程
+
+**快速开始**
+- [多模态理解：新手入门体验 [示例：InternVL2模型]](paddlemix/examples/internvl2/README.md)
+- [多模态生成：零基础上手指南 [示例：Stable Diffusion模型]](ppdiffusers/examples/stable_diffusion/README.md)
+- [跨模态任务流水线：快速开始](applications/README.md/#快速开始)
+
+**实操演练&范例**
+- [LLaVA模型：从训练到推理的全流程实践](https://aistudio.baidu.com/projectdetail/7917712)
+- [SDXL应用：打造专属奥运海报生成器](https://aistudio.baidu.com/projectdetail/8251202)
+- [飞桨PaddleMIX跨模态AI应用：项目分类汇总](./paddlemix_applications.md)
+
+**多硬件使用**
+- 昇腾910B支持的模型列表和使用方式，可以参考[昇腾硬件使用](./docs/hardware_support/ascend_usage.md)
+
+
+**数据准备&训练微调**
+- [模型训练与微调技巧](paddlemix/tools/README.md)
+
+**推理部署**
+- [部署指南：从开发到生产环境](deploy/README.md)
+
+
+## 📱模型库
+<table align="center">
+  <tbody>
+    <tr align="center" valign="center">
+      <td>
+        <b>多模态理解</b>
+      </td>
+      <td>
+        <b>多模态生成</b>
+      </td>
+      <td>
+        <b>多模态大一统</b>
+      </td>
+    </tr>
+    <tr valign="top">
+      <td>
+        <ul>
+        </ul>
+          <li><b>图文预训练</b></li>
+        <ul>
+            <li><a href="paddlemix/examples/clip">CLIP</a></li>
+            <li><a href="paddlemix/examples/evaclip">EVA-CLIP</a></li>
+            <li><a href="paddlemix/examples/llava">LLaVA-1.5</a></li>
+            <li><a href="paddlemix/examples/llava">LLaVA-1.6</a></li>
+            <li><a href="paddlemix/examples/llava">LLaVA-NeXT</a></li>
+            <li><a href="paddlemix/examples/llava_onevision">LLaVA-onevision</a></li>
+            <li><a href="paddlemix/examples/llava_onevision">Aquila-VL-2B-llava-qwen</a></li>
+            <li><a href="paddlemix/examples/llava_critic">LLaVA-Critic</a></li>
+            <li><a href="paddlemix/examples/llava_denseconnector">LLaVA-DenseConnector</a></li>
+            <li><a href="paddlemix/examples/qwen_vl">Qwen-VL</a></li>
+            <li><a href="paddlemix/examples/qwen2_vl">Qwen2-VL</a></li>
+            <li><a href="paddlemix/examples/internvl2">InternVL2</a></li>
+            <li><a href="paddlemix/examples/minimonkey">Mini-Monkey</a></li>
+            <li><a href="paddlemix/examples/coca">CoCa</a></li>
+            <li><a href="paddlemix/examples/blip2">BLIP-2</a></li>
+            <li><a href="paddlemix/examples/minigpt4">miniGPT-4</a></li>
+            <li><a href="paddlemix/examples/visualglm">VIsualGLM</a></li>
+            <li><a href="paddlemix/examples/cogvlm">CogVLM && CogAgent</a></li>
+            <li><a href="paddlemix/examples/internlm_xcomposer2">InternLM-XComposer2</a></li>
+      </ul>
+      </ul>
+          <li><b>开放世界视觉模型</b></li>
+        <ul>
+            <li><a href="paddlemix/examples/groundingdino">Grounding DINO</a></li>
+            <li><a href="paddlemix/examples/sam">SAM</a></li>
+            <li><a href="paddlemix/examples/YOLO-World">YOLO-World</a></li>
+      </ul>
+      </ul>
+          <li><b>更多模态预训练模型</b></li>
+        <ul>
+            <li><a href="paddlemix/examples/imagebind">ImageBind</a></li>
+      </ul>
+      </ul>
+        <li><b>数据分析</b></li>
+      <ul>
+          <li><a href="./paddlemix/datacopilot/example/pp_inscaptagger/">PP-InsCapTagger</a></li>
+      </ul>
+      </td>
+      <td>
+        <ul>
+        </ul>
+          <li><b>文生图</b></li>
+        <ul>
+           <li><a href="ppdiffusers/examples/stable_diffusion">Stable Diffusion</a></li>
+           <li><a href="ppdiffusers/ppdiffusers/pipelines/stable_diffusion_xl">SDXL</a></li>
+           <li><a href="ppdiffusers/examples/dreambooth/README_sd3.md">Stable Diffusion 3 (SD3)</a></li>
+            <li><a href="ppdiffusers/examples/controlnet">ControlNet</a></li>
+            <li><a href="ppdiffusers/examples/t2i-adapter">T2I-Adapter</a></li>
+            <li><a href="ppdiffusers/examples/text_to_image_laion400m">LDM</a></li>
+            <li><a href="ppdiffusers/ppdiffusers/pipelines/consistency_models">Consistency Models</a></li>
+            <li><a href="ppdiffusers/ppdiffusers/pipelines/deepfloyd_if">DeepFloyd IF</a></li>
+            <li><a href="ppdiffusers/ppdiffusers/pipelines/shap_e">Shap-E</a></li>
+            <li><a href="ppdiffusers/examples/kandinsky2_2">Kandinsky-2</a></li>
+            <li><a href="ppdiffusers/ppdiffusers/pipelines/wuerstchen">Würstchen</a></li>
+            <li><a href="ppdiffusers/ppdiffusers/pipelines/hotshot_xl">Hotshot-XL</a></li>
+            <li><a href="ppdiffusers/ppdiffusers/pipelines/latent_consistency_models">LCMs</a></li>
+            <li><a href="ppdiffusers/ppdiffusers/pipelines/unidiffuser">Unidiffuser</a></li>
+            <li><a href="ppdiffusers/examples/class_conditional_image_generation/DiT">DiT</a></li>
+            <li><a href="ppdiffusers/examples/HunyuanDiT">HunyuanDiT</a></li>
+        </ul>
+        </ul>
+          <li><b>文生视频</b></li>
+        <ul>
+           <li><a href="ppdiffusers/examples/text_to_video_lvdm">LVDM</a></li>
+           <li><a href="ppdiffusers/examples/stable_video_diffusion">SVD</a></li>
+           <li><a href="ppdiffusers/examples/AnimateAnyone">AnimateAnyone</a></li>
+           <li><a href="ppdiffusers/examples/Open-Sora">OpenSora</a></li>
+           <li><a href="ppdiffusers/ppdiffusers/pipelines/animatediff">AnimateDiff</a></li>
+           <li><a href="ppdiffusers/ppdiffusers/pipelines/text_to_video_synthesis">zeroscope_v2_XL</a></li>
+           <li><a href="ppdiffusers/examples/cogvideo">CogVideoX</a></li>
+        </ul>
+        </ul>
+          <li><b>音频生成</b></li>
+        <ul>
+           <li><a href="ppdiffusers/ppdiffusers/pipelines/audioldm">AudioLDM</a></li>
+           <li><a href="ppdiffusers/ppdiffusers/pipelines/audioldm2">AudioLDM2</a></li>
+        </ul>
+      </td>
+      <td>
+        <ul>
+        </ul>
+          <li><b>统一多模态模型</b></li>
+        <ul>
+          <li><a href="paddlemix/examples/janus">Janus</a></li>
+        </ul>
+      </td>
+    </tr>
+  </tbody>
+</table>
+
+更多模型能力，可参考[模型能力矩阵](./paddlemix/examples/README.md)
+
+
+## 🏆特色模型|工具
+
+### 💎跨模态任务流水线AppFlow
+<details>
+<summary><b> 简介(点击展开)</b></summary>
+
+AppFlow作为PaddleMIX的跨模态应用任务流水线，具备强大的功能与易用性。通过接入LLaVA、Stable Diffusion等前沿算法，AppFlow已全面覆盖图像、文本、音频、视频等多种模态，并通过流水线式的灵活组合，构建了10余种多模态应用，涵盖图文生成、文本视频生成、文本音频生成、图像理解等多个方面，为用户提供丰富的demo示例。AppFlow的特色在于其一键预测功能，用户无需繁琐训练与大量编码，��需简单命令即可完成模型推理，极大地降低了使用门槛。同时，AppFlow充分利用飞桨框架动静统一优势，用户只需设置简单参数，即可自动完成模型的动转静导出及高性能推理，提高工作效率并优化模型性能，实现一站式应用部署。
+
+`传送门`：[应用文档示例](applications/README.md/#快速开始)。
+
+</details>
+
+### 💎多模态数据处理工具箱DataCopilot
+<details>
+<summary><b> 简介(点击展开)</b></summary>
+
+在真实的应用场景有大量使用专有数据微调多模态大模型来提升模型效果的需求，此过程中数据要素成为核心。基于此PaddleMIX提供了数据处理和分析的工具DataCopilot，使开发者可在PaddleMIX套件完成端到端的开发体验。
+
+PP-InsCapTagger(Instance Capability Tagger) 是 DataCopilot 基于 PaddleMIX 实现的数据集能力标签模型，用于为多模态数据实例能力打标，通过实例能力分布对数据集进行优化，可以提高模型训练效率，为数据集分析和评价提供了一种高效的方案。 结合模型推理打标结果对LLaVA SFT数据集进行优化，可以**提高LLaVA模型SFT阶段50%的训练效率。**
+
+`传送门`：[应用文档示例](paddlemix/datacopilot/readme.md)。
+
+</details>
+
+<details>
+<summary><b> PP-InsCapTagger(点击展开)</b></summary>
+
+| Model                           | ScienceQA                               | TextVQA                                | VQAv2                                  | GQA                                    | MMMU                                   | MME                                     |
+|----------------------------------|-----------------------------------------|----------------------------------------|----------------------------------------|----------------------------------------|----------------------------------------|-----------------------------------------|
+| llava-1.5-7b (origin)            | 66.8                                    | 58.2                                   | 78.5                                   | 62                                     | -                                      | -                                       |
+| llava-1.5-7b (rerun)             | 69.01                                   | 57.6                                   | 79                                     | 62.95                                  | 36.89                                  | 1521<br>323                             |
+| llava-1.5-7b (random 50%)        | 67.31                                   | 55.6                                   | 76.89                                  | 61.01                                  | 34.67                                  | 1421<br>286                             |
+| **llava-1.5-7b (our 50%)**       | **70.24** *(+2.93)*                     | **57.12** *(+1.52)*                    | **78.32** *(+1.43)*                    | **62.14** *(+1.13)*                    | **37.11** *(+2.44)*                    | **1476** *(+55)*<br>**338** *(+52)*    |
+
+
+`传送门`：[应用文档示例](paddlemix/datacopilot/example/pp_inscaptagger/readme.md)。
+</details>
+
+
+## 🤔FAQ
+关于我们项目的一些常见问题解答，请参考[FAQ](docs/FAQ.md)。如果您的问题没有得到解答，请随时在[Issues](https://github.com/PaddlePaddle/PaddleMIX/issues)中提出
+
+
+## 📝许可证书
+
+本项目的发布受[Apache 2.0 license](LICENSE)许可认证。
+
+## 📌社区交流
+
+- 微信扫描二维码并填写问卷，即可加入交流群与众多社区开发者以及官方团队深度交流。
+<div align="center">
+    <img src="https://github.com/user-attachments/assets/ecf292da-9ac6-41cb-84b6-df726ef4522d" width="300" height="300" />
+</div>

PaddleMIX/paddlemix_applications.md

@@ -0,0 +1,244 @@
+# 🎨 飞桨PaddleMIX项目分类汇总
+
+## 🎯 一、多模态创作类
+
+1. [PaddleMIX开学季海报生成](https://aistudio.baidu.com/projectdetail/8240734)
+    - 🌇 类型：文生图
+    - 🎓 功能：开学季和教师节主题海报生成
+    - 🔥 热度：🔥🔥🔥
+
+2. [PaddleMIX黑神话悟空海报生成](https://aistudio.baidu.com/projectdetail/8555550)
+    - 🎨 类型：文生图
+    - 🎮 功能：黑神话悟空主题海报生成
+    - 🔥 热度：🔥🔥🔥
+
+3. [AI智能节日海报](https://aistudio.baidu.com/projectdetail/8293232)
+    - 🎨 类型：文生图
+    - 🎉 功能：文心大模型与PaddleMIX驱动的AI智能海报生成
+    - 🔥 热度：🔥🔥
+
+4. [AI建筑设计助手](https://aistudio.baidu.com/projectdetail/8288584)
+    - 🏗️ 类型：图像渲染
+    - 🎯 功能：建筑设计图快速渲染解决方案
+    - 🔥 热度：🔥🔥
+
+5. [PaddleMIX教师节文图生成器](https://aistudio.baidu.com/projectdetail/8530911)
+    - 📚 类型：文生图
+    - 🎓 功能：教师节主题创意生成
+    - 🔥 热度：🔥🔥
+
+6. [文本+音乐生成图片](https://aistudio.baidu.com/projectdetail/8314111)
+    - 🎵 类型：多模态生成
+    - 🖼️ 功能：基于文本和音乐的图像生成
+    - 🔥 热度：🔥🔥
+
+7. [个性化GIF生成](https://aistudio.baidu.com/projectdetail/8321341)
+    - 🌠 类型：文生图
+    - 🍡 功能：基于Hotshot-XL模型的GIF生成
+    - 🔥 热度：🔥🔥
+
+8. [绿心智绘：环保创意海报](https://aistudio.baidu.com/projectdetail/8362605)
+    - 🌻 类型：文生图
+    - 🍀 功能：环保创意海报生成
+    - 🔥 热度：🔥🔥
+
+9. [墨韵融生:多模态水墨创艺](https://aistudio.baidu.com/projectdetail/8515323)
+    - 🌉 类型：文生图
+    - 🎨 功能：水墨画作品生成
+    - 🔥 热度：🔥
+
+10. [PaddleMIX的PPT生成器](https://aistudio.baidu.com/projectdetail/8545186)
+    - 💻 类型：幻灯片生成
+    - 💡 功能：内容生成、图片设计和PPT文件渲染
+    - 🔥 热度：🔥
+
+11. [那兔海报](https://aistudio.baidu.com/projectdetail/8331103)
+    - 🐰 类型：文生图
+    - 🔅 功能：那兔海报生成
+    - 🔥 热度：🔥
+
+## 🎬 二、视频创作类
+12. [PaddleMIX AnimateAnyone模型应用](https://aistudio.baidu.com/projectdetail/8320311)
+    - 🎥 类型：图生视频
+    - 🎭 功能：依据图片创作视频
+    - 🔥 热度：🔥🔥
+
+13. [OpenSora模型应用](https://aistudio.baidu.com/projectdetail/8534189)
+    - 🎦 类型：文生视频/图生视频
+    - 🎪 功能：多模态视频生成
+    - 🔥 热度：🔥🔥
+
+14. [openSora和SVD视频渲染引擎](https://aistudio.baidu.com/projectdetail/8535627)
+    - 🎮 类型：图像转视频
+    - 🎨 功能：图像转换视频渲染
+    - 🔥 热度：🔥🔥
+
+## 🎵 三、音乐与音频类
+15. [PaddleMIX音乐快速生成器](https://aistudio.baidu.com/projectdetail/8556362)
+    - 🎧 类型：音乐生成
+    - 🎺 功能：基于文本的音乐生成
+    - 🔥 热度：🔥🔥🔥
+
+16. [PaddleMIX照片配音生成](https://aistudio.baidu.com/projectdetail/7454890)
+    - 🎸 类型：音乐生成
+    - 🎼 功能：照片的配音生成
+    - 🔥 热度：🔥🔥🔥
+
+17. [智能视频配乐](https://aistudio.baidu.com/projectdetail/8351571)
+    - 🎼 类型：音频生成
+    - 🎹 功能：视频氛围分析与背景音乐生成
+    - 🔥 热度：🔥🔥
+
+18. [国庆节音乐贺卡](https://aistudio.baidu.com/projectdetail/8536882)
+    - 🎸 类型：音乐生成
+    - 🎺 功能：国庆专属电子音乐贺卡
+    - 🔥 热度：🔥🔥
+
+19. [月圆之夜：音乐与中秋的邂逅](https://aistudio.baidu.com/projectdetail/8311291)
+    - 🌕 类型：音乐生成
+    - 🎶 功能：中秋节主题音乐创作
+    - 🔥 热度：🔥🔥
+
+20. [音画交响：图片配乐新纪元](https://aistudio.baidu.com/projectdetail/8398422)
+    - 🎻 类型：音乐生成
+    - 🎺 功能：照片的配乐生成
+    - 🔥 热度：🔥
+
+21. [国庆庆典：音乐与国庆辉煌交响](https://aistudio.baidu.com/projectdetail/8356701)
+    - 🎹 类型：音乐生成
+    - 🎺 功能：基于文本和照片的配乐生成
+    - 🔥 热度：🔥
+
+## 🎊 四、节日主题创作
+22. [国庆写真生成器](https://aistudio.baidu.com/projectdetail/8334575)
+    - 🎀 类型：图像生成
+    - 👗 功能：专属国庆节写真生成
+    - 🔥 热度：🔥🔥🔥 
+
+23. [中秋民俗舞火龙](https://aistudio.baidu.com/projectdetail/8318522)
+    - 🐉 类型：图像生成
+    - 🔥 功能：舞火龙庆典图绘制
+    - 🔥 热度：🔥🔥🔥 
+
+24. [兔绘爱国情：童心海报梦工厂](https://aistudio.baidu.com/projectdetail/8288021)
+    - 🐰 类型：图像生成
+    - 🎨 功能：爱国主题童趣海报
+    - 🔥 热度：🔥🔥
+
+25. [创艺·幻绘月饼艺术图](https://aistudio.baidu.com/projectdetail/8540279)
+    - 🥮 类型：图像生成
+    - 🎨 功能：月饼艺术创作
+    - 🔥 热度：🔥🔥
+
+26. [PaddleMIX中秋贺卡生成器](https://aistudio.baidu.com/projectdetail/8541174)
+    - 🌙 类型：图像生成
+    - ✉️ 功能：中秋节贺卡定制
+    - 🔥 热度：🔥🔥
+
+27. [国庆英雄颂](https://aistudio.baidu.com/projectdetail/8352450)
+    - 🇨🇳 类型：多模态创作
+    - 📣 功能：历史的英雄故事以生动的声音和视觉形式呈现
+    - 🔥 热度：🔥🔥
+
+28. [国庆烟火绘梦](https://aistudio.baidu.com/projectdetail/8352778)
+    - 🎆 类型：文生图
+    - 🎇 功能：数字烟花的生成
+    - 🔥 热度：🔥🔥
+
+29. [绘梦国庆-缤纷国庆贺卡](https://aistudio.baidu.com/projectdetail/8542328)
+    - 🎉 类型：图像生成
+    - 🎊 功能：个性化国庆贺卡创作
+    - 🔥 热度：🔥
+
+30. [感恩有你-感恩节海报设计](https://aistudio.baidu.com/projectdetail/8500805)
+    - 🍔 类型：图像生成
+    - 🍖 功能：感恩节海报生成
+    - 🔥 热度：🔥
+
+31. [魔境创想-万圣节贺卡生成](https://aistudio.baidu.com/projectdetail/8500805)
+    - 🎃 类型：图像生成
+    - 👻 功能：个性化的万圣节贺卡生成
+    - 🔥 热度：🔥
+
+## 🤖 五、智能助手类
+32. [炉石传说卡牌设计助手](https://aistudio.baidu.com/projectdetail/8543362)
+    - 🎮 类型：多模态设计
+    - 🃏 功能：基于ErnieSDK的游戏卡牌设计
+    - 🔥 热度：🔥🔥
+
+33. [智能宠物识别与养护助手](https://aistudio.baidu.com/projectdetail/8299163)
+    - 🐾 类型：图像识别/分析
+    - 🐱 功能：基于QwenVL的宠物护理
+    - 🔥 热度：🔥🔥
+
+## 💡 六、创新应用类
+34. [多模态数字艺术创生](https://aistudio.baidu.com/projectdetail/8382636)
+    - 🎨 类型：艺术创作
+    - 🖼️ 功能：数字艺术创作与鉴赏
+    - 🔥 热度：🔥🔥
+
+35. [教学辅助多模态图文生成](https://aistudio.baidu.com/projectdetail/8372762)
+    - 📚 类型：教育应用
+    - 🎓 功能：教学辅助创作工具
+    - 🔥 热度：🔥🔥
+
+36. ["海参纪念币"文创设计](https://aistudio.baidu.com/projectdetail/8386127)
+    - 💰 类型：文创设计
+    - 🎨 功能：纪念币艺术设计
+    - 🔥 热度：🔥🔥
+
+37. [艺术风格迁移](https://aistudio.baidu.com/projectdetail/8549441)
+    - 🎨 类型：风格迁移
+    - 🖼️ 功能：多模态艺术风格迁移
+    - 🔥 热度：🔥🔥
+
+38. [爱国卡通角色故事创作](https://aistudio.baidu.com/projectdetail/8356273)
+    - 📚 类型：多模态创作
+    - 🎭 功能：基于ERNIE SDK的角色故事创作
+    - 🔥 热度：🔥
+
+39. [云顶天宫：盗墓笔记场景重现](https://aistudio.baidu.com/projectdetail/8486711)
+    - 📚 类型：多模态创作
+    - 🎭 功能：将小说中场景深度还原
+    - 🔥 热度：🔥
+
+40. [绘梦长空·风筝韵事](https://aistudio.baidu.com/projectdetail/8404865)
+    - 🎉 类型：图像理解
+    - 📚 功能：讲述中国传统风筝背后的动人故事与制作工艺
+    - 🔥 热度：🔥 
+
+## 七、其他
+41. [基于PaddleMIX的数据集行为标签分类器训推实例](https://aistudio.baidu.com/projectdetail/7917712)
+    - 🎓 类型：图像理解
+    - 🌋 功能：基于LLaVA的行为分类
+    - 🔥 热度：🔥🔥🔥
+    
+42. [PaddleMIX入门AIGC应用](https://aistudio.baidu.com/projectdetail/7583868)
+    - 📟 类型：多模态生成
+    - 📷 功能：图像生成、音频生成、图像变换、图像超分辨率
+    - 🔥 热度：🔥🔥
+
+43. [PaddleMIX DiT高性能推理实战](https://aistudio.baidu.com/projectdetail/8261962)
+    - 🎨 类型：图像生成
+    - 📚 功能：基于DiT结构的图像生成
+    - 🔥 热度：🔥🔥
+
+44. [PaddleMIX玩转Stable Diffusion 3](https://aistudio.baidu.com/projectdetail/8494762)
+    - 🔔 类型：图像生成
+    - 🎮 功能：基于Stable Diffusion 3的图像生成
+    - 🔥 热度：🔥🔥
+
+45. [基于PaddleMIX的数据集行为标签分类器训推实例](https://aistudio.baidu.com/projectdetail/8229498)
+    - 🎈 类型：图像理解
+    - 💎 功能：LLaVA模型训练与使用
+    - 🔥 热度：🔥🔥
+
+46. [PaddleMIX入门视频生成](https://aistudio.baidu.com/projectdetail/8221041)
+    - 🎥 类型：视频生成
+    - 📺 功能：基于SVD和OpenSora的视频生成
+    - 🔥 热度：🔥
+
+47. [基于PaddleMIX的高速文生图推理](https://aistudio.baidu.com/projectdetail/8232803)
+    - 🎓 类型：图像生成
+    - 🎨 功能：LCM模型的快速图像生成
+    - 🔥 热度：🔥

VILA/Dockerfile

@@ -0,0 +1,18 @@
+FROM nvcr.io/nvidia/pytorch:24.06-py3
+
+WORKDIR /app
+
+RUN curl https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -o ~/miniconda.sh \
+    && sh ~/miniconda.sh -b -p /opt/conda \
+    && rm ~/miniconda.sh
+
+ENV PATH /opt/conda/bin:$PATH
+COPY pyproject.toml pyproject.toml
+COPY llava llava
+
+COPY environment_setup.sh environment_setup.sh
+RUN bash environment_setup.sh vila
+
+
+COPY server.py server.py
+CMD ["conda", "run", "-n", "vila", "--no-capture-output", "python", "-u", "-W", "ignore", "server.py"]

VILA/server.py

@@ -0,0 +1,254 @@
+import argparse
+import base64
+import os
+import re
+import time
+import uuid
+from contextlib import asynccontextmanager
+from io import BytesIO
+from typing import List, Literal, Optional, Union, get_args
+
+import requests
+import torch
+import uvicorn
+from fastapi import FastAPI
+from fastapi.responses import JSONResponse
+from PIL import Image as PILImage
+from PIL.Image import Image
+from pydantic import BaseModel
+
+from llava.constants import (
+    DEFAULT_IM_END_TOKEN,
+    DEFAULT_IM_START_TOKEN,
+    DEFAULT_IMAGE_TOKEN,
+    IMAGE_PLACEHOLDER,
+    IMAGE_TOKEN_INDEX,
+)
+from llava.conversation import SeparatorStyle, conv_templates
+from llava.mm_utils import KeywordsStoppingCriteria, get_model_name_from_path, process_images, tokenizer_image_token
+from llava.model.builder import load_pretrained_model
+from llava.utils import disable_torch_init
+
+
+class TextContent(BaseModel):
+    type: Literal["text"]
+    text: str
+
+
+class ImageURL(BaseModel):
+    url: str
+
+
+class ImageContent(BaseModel):
+    type: Literal["image_url"]
+    image_url: ImageURL
+
+
+IMAGE_CONTENT_BASE64_REGEX = re.compile(r"^data:image/(png|jpe?g);base64,(.*)$")
+
+
+class ChatMessage(BaseModel):
+    role: Literal["user", "assistant"]
+    content: Union[str, List[Union[TextContent, ImageContent]]]
+
+
+class ChatCompletionRequest(BaseModel):
+    model: Literal[
+        "VILA1.5-3B",
+        "VILA1.5-3B-AWQ",
+        "VILA1.5-3B-S2",
+        "VILA1.5-3B-S2-AWQ",
+        "Llama-3-VILA1.5-8B",
+        "Llama-3-VILA1.5-8B-AWQ",
+        "VILA1.5-13B",
+        "VILA1.5-13B-AWQ",
+        "VILA1.5-40B",
+        "VILA1.5-40B-AWQ",
+    ]
+    messages: List[ChatMessage]
+    max_tokens: Optional[int] = 512
+    top_p: Optional[float] = 0.9
+    temperature: Optional[float] = 0.2
+    stream: Optional[bool] = False
+    use_cache: Optional[bool] = True
+    num_beams: Optional[int] = 1
+
+
+model = None
+model_name = None
+tokenizer = None
+image_processor = None
+context_len = None
+
+
+def load_image(image_url: str) -> Image:
+    if image_url.startswith("http") or image_url.startswith("https"):
+        response = requests.get(image_url)
+        image = PILImage.open(BytesIO(response.content)).convert("RGB")
+    else:
+        match_results = IMAGE_CONTENT_BASE64_REGEX.match(image_url)
+        if match_results is None:
+            raise ValueError(f"Invalid image url: {image_url}")
+        image_base64 = match_results.groups()[1]
+        image = PILImage.open(BytesIO(base64.b64decode(image_base64))).convert("RGB")
+    return image
+
+
+def get_literal_values(cls, field_name: str):
+    field_type = cls.__annotations__.get(field_name)
+    if field_type is None:
+        raise ValueError(f"{field_name} is not a valid field name")
+    if hasattr(field_type, "__origin__") and field_type.__origin__ is Literal:
+        return get_args(field_type)
+    raise ValueError(f"{field_name} is not a Literal type")
+
+
+VILA_MODELS = get_literal_values(ChatCompletionRequest, "model")
+
+
+def normalize_image_tags(qs: str) -> str:
+    image_token_se = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN
+    if IMAGE_PLACEHOLDER in qs:
+        if model.config.mm_use_im_start_end:
+            qs = re.sub(IMAGE_PLACEHOLDER, image_token_se, qs)
+        else:
+            qs = re.sub(IMAGE_PLACEHOLDER, DEFAULT_IMAGE_TOKEN, qs)
+
+    if DEFAULT_IMAGE_TOKEN not in qs:
+        print("No image was found in input messages. Continuing with text only prompt.")
+    return qs
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    global model, model_name, tokenizer, image_processor, context_len
+    disable_torch_init()
+    model_path = app.args.model_path
+    model_name = get_model_name_from_path(model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, model_name, None)
+    print(f"Model {model_name} loaded successfully. Context length: {context_len}")
+    yield
+
+
+app = FastAPI(lifespan=lifespan)
+
+
+# Load model upon startup
+@app.post("/chat/completions")
+async def chat_completions(request: ChatCompletionRequest):
+    try:
+        global model, tokenizer, image_processor, context_len
+
+        if request.model != model_name:
+            raise ValueError(
+                f"The endpoint is configured to use the model {model_name}, "
+                f"but the request model is {request.model}"
+            )
+        max_tokens = request.max_tokens
+        temperature = request.temperature
+        top_p = request.top_p
+        use_cache = request.use_cache
+        num_beams = request.num_beams
+
+        messages = request.messages
+        conv_mode = app.args.conv_mode
+
+        images = []
+
+        conv = conv_templates[conv_mode].copy()
+        user_role = conv.roles[0]
+        assistant_role = conv.roles[1]
+
+        for message in messages:
+            if message.role == "user":
+                prompt = ""
+
+                if isinstance(message.content, str):
+                    prompt += message.content
+                if isinstance(message.content, list):
+                    for content in message.content:
+                        if content.type == "text":
+                            prompt += content.text
+                        if content.type == "image_url":
+                            image = load_image(content.image_url.url)
+                            images.append(image)
+                            prompt += IMAGE_PLACEHOLDER
+
+                normalized_prompt = normalize_image_tags(prompt)
+                conv.append_message(user_role, normalized_prompt)
+            if message.role == "assistant":
+                prompt = message.content
+                conv.append_message(assistant_role, prompt)
+
+        prompt_text = conv.get_prompt()
+        print("Prompt input: ", prompt_text)
+
+        # support generation with text only inputs
+        if len(images) == 0:
+            images_input = None
+        else:
+            images_tensor = process_images(images, image_processor, model.config).to(model.device, dtype=torch.float16)
+            images_input = [images_tensor]
+
+        input_ids = (
+            tokenizer_image_token(prompt_text, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt")
+            .unsqueeze(0)
+            .to(model.device)
+        )
+
+        stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
+        keywords = [stop_str]
+        stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
+
+        with torch.inference_mode():
+            output_ids = model.generate(
+                input_ids,
+                images=images_input,
+                do_sample=True if temperature > 0 else False,
+                temperature=temperature,
+                top_p=top_p,
+                num_beams=num_beams,
+                max_new_tokens=max_tokens,
+                use_cache=use_cache,
+                stopping_criteria=[stopping_criteria],
+            )
+
+        outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0]
+        outputs = outputs.strip()
+        if outputs.endswith(stop_str):
+            outputs = outputs[: -len(stop_str)]
+        outputs = outputs.strip()
+        print("\nAssistant: ", outputs)
+
+        resp_content = [TextContent(type="text", text=outputs)]
+        return {
+            "id": uuid.uuid4().hex,
+            "object": "chat.completion",
+            "created": time.time(),
+            "model": request.model,
+            "choices": [{"message": ChatMessage(role="assistant", content=resp_content)}],
+        }
+    except Exception as e:
+        return JSONResponse(
+            status_code=500,
+            content={"error": str(e)},
+        )
+
+
+if __name__ == "__main__":
+
+    host = os.getenv("VILA_HOST", "0.0.0.0")
+    port = os.getenv("VILA_PORT", 8000)
+    model_path = os.getenv("VILA_MODEL_PATH", "Efficient-Large-Model/VILA1.5-3B")
+    conv_mode = os.getenv("VILA_CONV_MODE", "vicuna_v1")
+    workers = os.getenv("VILA_WORKERS", 1)
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", type=str, default=host)
+    parser.add_argument("--port", type=int, default=port)
+    parser.add_argument("--model-path", type=str, default=model_path)
+    parser.add_argument("--conv-mode", type=str, default=conv_mode)
+    parser.add_argument("--workers", type=int, default=workers)
+    app.args = parser.parse_args()
+
+    uvicorn.run(app, host=host, port=port, workers=workers)

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/commands/env.py

@@ -0,0 +1,82 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import platform
+from argparse import ArgumentParser
+
+from ..utils import is_paddle_available, is_paddlenlp_available, is_ppxformers_available
+from ..version import VERSION as version
+from . import BasePPDiffusersCLICommand
+
+
+def info_command_factory(_):
+    return EnvironmentCommand()
+
+
+class EnvironmentCommand(BasePPDiffusersCLICommand):
+    @staticmethod
+    def register_subcommand(parser: ArgumentParser):
+        download_parser = parser.add_parser("env")
+        download_parser.set_defaults(func=info_command_factory)
+
+    def run(self):
+        import huggingface_hub
+
+        hub_version = huggingface_hub.__version__
+
+        pd_version = "not installed"
+        pd_cuda_available = "NA"
+        if is_paddle_available():
+            import paddle
+
+            pd_version = paddle.__version__
+            pd_cuda_available = paddle.device.is_compiled_with_cuda()
+
+        paddlenlp_version = "not installed"
+        if is_paddlenlp_available:
+            import paddlenlp
+
+            paddlenlp_version = paddlenlp.__version__
+
+        from ppdiffusers.accelerate import __version__ as accelerate_version
+        from ppdiffusers.peft import __version__ as peft_version
+
+        xformers_commit_id = "not installed"
+        if is_ppxformers_available():
+            import paddle
+
+            xformers_commit_id = paddle.__git_commit__
+
+        info = {
+            "`ppdiffusers` version": version,
+            "Platform": platform.platform(),
+            "Python version": platform.python_version(),
+            "Paddle version (GPU?)": f"{pd_version} ({pd_cuda_available})",
+            "Huggingface_hub version": hub_version,
+            "PaddleNLP version": paddlenlp_version,
+            "PP-Accelerate version": accelerate_version,
+            "PP-Peft version": peft_version,
+            "PP-xFormers commit id": xformers_commit_id,
+            "Using GPU in script?": "<fill in>",
+            "Using distributed or parallel set-up in script?": "<fill in>",
+        }
+
+        print("\nCopy-and-paste the text below in your GitHub issue and FILL OUT the two last points.\n")
+        print(self.format_dict(info))
+
+        return info
+
+    @staticmethod
+    def format_dict(d):
+        return "\n".join([f"- {prop}: {val}" for prop, val in d.items()]) + "\n"

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/experimental/README.md

@@ -0,0 +1,5 @@
+# 🧨 Diffusers Experimental
+
+We are adding experimental code to support novel applications and usages of the Diffusers library.
+Currently, the following experiments are supported:
+* Reinforcement learning via an implementation of the [Diffuser](https://arxiv.org/abs/2205.09991) model.

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/experimental/__init__.py

@@ -0,0 +1,15 @@
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .rl import ValueGuidedRLPipeline

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/loaders/__init__.py

@@ -0,0 +1,54 @@
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+from ..utils import PPDIFFUSERS_SLOW_IMPORT, _LazyModule
+from ..utils.import_utils import is_paddle_available, is_paddlenlp_available
+
+_import_structure = {}
+
+if is_paddle_available():
+    _import_structure["single_file"] = ["FromOriginalControlnetMixin", "FromOriginalVAEMixin"]
+    _import_structure["unet"] = ["UNet2DConditionLoadersMixin"]
+    _import_structure["utils"] = ["AttnProcsLayers"]
+
+    if is_paddlenlp_available():
+        _import_structure["single_file"].extend(["FromSingleFileMixin", "FromCkptMixin"])
+        _import_structure["lora"] = ["LoraLoaderMixin", "StableDiffusionXLLoraLoaderMixin", "SD3LoraLoaderMixin"]
+        _import_structure["textual_inversion"] = ["TextualInversionLoaderMixin"]
+        _import_structure["ip_adapter"] = ["IPAdapterMixin"]
+        # NOTE: this will removed in the future
+        _import_structure["deprecate"] = ["text_encoder_lora_state_dict", "text_encoder_attn_modules"]
+
+if TYPE_CHECKING or PPDIFFUSERS_SLOW_IMPORT:
+    if is_paddle_available():
+        from .single_file import FromOriginalControlnetMixin, FromOriginalVAEMixin
+        from .unet import UNet2DConditionLoadersMixin
+        from .utils import AttnProcsLayers
+
+        if is_paddlenlp_available():
+            # NOTE: this will removed in the future
+            from .deprecate import (
+                text_encoder_attn_modules,
+                text_encoder_lora_state_dict,
+            )
+            from .ip_adapter import IPAdapterMixin
+            from .lora import LoraLoaderMixin, SD3LoraLoaderMixin, StableDiffusionXLLoraLoaderMixin
+            from .single_file import FromCkptMixin, FromSingleFileMixin
+            from .textual_inversion import TextualInversionLoaderMixin
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/loaders/deprecate.py

@@ -0,0 +1,63 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ..utils import deprecate
+from ..utils.import_utils import is_paddlenlp_available
+
+
+def text_encoder_lora_state_dict(text_encoder):
+    deprecate(
+        "text_encoder_load_state_dict in `models`",
+        "0.45.0",
+        "`text_encoder_lora_state_dict` is deprecated and will be removed in 0.27.0. Make sure to retrieve the weights using `get_peft_model`. See https://huggingface.co/docs/peft/v0.6.2/en/quicktour#peftmodel for more information.",
+    )
+    state_dict = {}
+
+    for name, module in text_encoder_attn_modules(text_encoder):
+        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
+
+        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
+
+        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
+
+        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
+
+    return state_dict
+
+
+if is_paddlenlp_available():
+
+    def text_encoder_attn_modules(text_encoder):
+        deprecate(
+            "text_encoder_attn_modules in `models`",
+            "0.45.0",
+            "`text_encoder_lora_state_dict` is deprecated and will be removed in 0.27.0. Make sure to retrieve the weights using `get_peft_model`. See https://huggingface.co/docs/peft/v0.6.2/en/quicktour#peftmodel for more information.",
+        )
+        from ppdiffusers.transformers import CLIPTextModel, CLIPTextModelWithProjection
+
+        attn_modules = []
+
+        if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
+            for i, layer in enumerate(text_encoder.text_model.encoder.layers):
+                name = f"text_model.encoder.layers.{i}.self_attn"
+                mod = layer.self_attn
+                attn_modules.append((name, mod))
+        else:
+            raise ValueError(f"do not know how to get attention modules for: {text_encoder.__class__.__name__}")
+
+        return attn_modules

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/loaders/ip_adapter.py

@@ -0,0 +1,196 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+from typing import Dict, Union
+
+import paddle
+from safetensors import safe_open
+
+from ..utils import (
+    DIFFUSERS_CACHE,
+    FROM_AISTUDIO,
+    FROM_DIFFUSERS,
+    FROM_HF_HUB,
+    HF_HUB_OFFLINE,
+    PPDIFFUSERS_CACHE,
+    _get_model_file,
+    is_paddlenlp_available,
+    logging,
+    smart_load,
+)
+
+if is_paddlenlp_available():
+    from ppdiffusers.transformers import (
+        CLIPImageProcessor,
+        CLIPVisionModelWithProjection,
+    )
+
+    from ..models.attention_processor import (
+        IPAdapterAttnProcessor,
+        IPAdapterAttnProcessor2_5,
+    )
+
+logger = logging.get_logger(__name__)
+
+
+class IPAdapterMixin:
+    """Mixin for handling IP Adapters."""
+
+    def load_ip_adapter(
+        self,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, paddle.Tensor]],
+        subfolder: str = "",
+        weight_name: str = None,
+        **kwargs,
+    ):
+        """
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                Can be either:
+
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            use_auth_token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+        """
+
+        # Load the main state dict first.
+        from_hf_hub = kwargs.pop("from_hf_hub", FROM_HF_HUB)
+        from_aistudio = kwargs.pop("from_aistudio", FROM_AISTUDIO)
+        cache_dir = kwargs.pop("cache_dir", None)
+        if cache_dir is None:
+            if from_aistudio:
+                cache_dir = None  # TODO, check aistudio cache
+            elif from_hf_hub:
+                cache_dir = DIFFUSERS_CACHE
+            else:
+                cache_dir = PPDIFFUSERS_CACHE
+        from_diffusers = kwargs.pop("from_diffusers", FROM_DIFFUSERS)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
+        revision = kwargs.pop("revision", None)
+        if subfolder is None:
+            subfolder = ""
+        user_agent = {
+            "file_type": "attn_procs_weights",
+            "framework": "pytorch" if from_diffusers else "paddle",
+        }
+        if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+            model_file = _get_model_file(
+                pretrained_model_name_or_path_or_dict,
+                weights_name=weight_name,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                resume_download=resume_download,
+                proxies=proxies,
+                local_files_only=local_files_only,
+                use_auth_token=use_auth_token,
+                revision=revision,
+                subfolder=subfolder,
+                user_agent=user_agent,
+                from_hf_hub=from_hf_hub,
+                from_aistudio=from_aistudio,
+            )
+            if weight_name.endswith(".safetensors"):
+                state_dict = {"image_proj": {}, "ip_adapter": {}}
+                with safe_open(model_file, framework="np") as f:
+                    metadata = f.metadata()
+                    if metadata is None:
+                        metadata = {}
+                    if metadata.get("format", "pt") not in ["pt", "pd", "np"]:
+                        raise OSError(
+                            f"The safetensors archive passed at {model_file} does not contain the valid metadata. Make sure "
+                            "you save your model with the `save_pretrained` method."
+                        )
+                    data_format = metadata.get("format", "pt")
+                    if data_format == "pt" and not from_diffusers:
+                        logger.warning(
+                            "Detect the weight is in diffusers format, but currently, `from_diffusers` is set to `False`. To proceed, we will change the value of `from_diffusers` to `True`!"
+                        )
+                        from_diffusers = True
+                    for key in f.keys():
+                        if key.startswith("image_proj."):
+                            state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
+                        elif key.startswith("ip_adapter."):
+                            state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+            else:
+                state_dict = smart_load(model_file, return_numpy=True, return_is_torch_weight=True)
+                is_torch_weight = state_dict.pop("is_torch_weight", False)
+                if not from_diffusers and is_torch_weight:
+                    logger.warning(
+                        "Detect the weight is in diffusers format, but currently, `from_diffusers` is set to `False`. To proceed, we will change the value of `from_diffusers` to `True`!"
+                    )
+                    from_diffusers = True
+        else:
+            state_dict = pretrained_model_name_or_path_or_dict
+
+        keys = list(state_dict.keys())
+        if sorted(keys) != ["image_proj", "ip_adapter"]:
+            raise ValueError("Required keys are (`image_proj` and `ip_adapter`) missing from the state dict.")
+
+        # load CLIP image encoer here if it has not been registered to the pipeline yet
+        if hasattr(self, "image_encoder") and getattr(self, "image_encoder", None) is None:
+            if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+                logger.info(f"loading image_encoder from {pretrained_model_name_or_path_or_dict}")
+                image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+                    pretrained_model_name_or_path_or_dict,
+                    subfolder=os.path.join(subfolder, "image_encoder"),
+                    from_hf_hub=from_hf_hub,
+                    from_aistudio=from_aistudio,
+                    # from_diffusers=from_diffusers, # we must disable this !
+                ).to(dtype=self.dtype)
+                self.image_encoder = image_encoder
+            else:
+                raise ValueError("`image_encoder` cannot be None when using IP Adapters.")
+
+        # create feature extractor if it has not been registered to the pipeline yet
+        if hasattr(self, "feature_extractor") and getattr(self, "feature_extractor", None) is None:
+            self.feature_extractor = CLIPImageProcessor()
+
+        # load ip-adapter into unet
+        self.unet._load_ip_adapter_weights(state_dict, from_diffusers=from_diffusers)
+
+    def set_ip_adapter_scale(self, scale):
+        for attn_processor in self.unet.attn_processors.values():
+            if isinstance(attn_processor, (IPAdapterAttnProcessor, IPAdapterAttnProcessor2_5)):
+                attn_processor.scale = scale

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/loaders/lora.py

@@ -0,0 +1,1871 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+from contextlib import nullcontext
+from functools import partial
+from typing import Callable, Dict, List, Optional, Union
+from pathlib import Path
+
+import numpy as np
+import paddle
+from huggingface_hub import model_info
+from packaging import version
+from paddle import nn
+
+from ..models.modeling_pytorch_paddle_utils import (
+    convert_paddle_state_dict_to_pytorch,
+    convert_pytorch_state_dict_to_paddle,
+)
+from ..models.modeling_utils import faster_set_state_dict
+from ..utils import (
+    DIFFUSERS_CACHE,
+    FROM_AISTUDIO,
+    FROM_DIFFUSERS,
+    FROM_HF_HUB,
+    HF_HUB_OFFLINE,
+    LOW_CPU_MEM_USAGE_DEFAULT,
+    PPDIFFUSERS_CACHE,
+    TO_DIFFUSERS,
+    USE_PEFT_BACKEND,
+    _get_model_file,
+    convert_state_dict_to_peft,
+    convert_state_dict_to_ppdiffusers,
+    convert_unet_state_dict_to_peft,
+    delete_adapter_layers,
+    deprecate,
+    get_adapter_name,
+    get_peft_kwargs,
+    is_paddlenlp_available,
+    is_safetensors_available,
+    is_torch_available,
+    logging,
+    recurse_remove_peft_layers,
+    scale_lora_layers,
+    set_adapter_layers,
+    set_weights_and_activate_adapters,
+)
+from ..version import VERSION as __version__
+from .lora_conversion_utils import (
+    _convert_kohya_lora_to_diffusers,
+    _maybe_map_sgm_blocks_to_diffusers,
+)
+
+if is_safetensors_available():
+    from safetensors.numpy import save_file as np_safe_save_file
+
+    if is_torch_available():
+        from safetensors.torch import save_file as torch_safe_save_file
+
+if is_torch_available():
+    import torch
+
+if is_paddlenlp_available():
+    from paddlenlp.transformers import PretrainedModel
+
+    from ..models.lora import (
+        PatchedLoraProjection,
+        text_encoder_attn_modules,
+        text_encoder_mlp_modules,
+    )
+
+
+logger = logging.get_logger(__name__)
+
+TEXT_ENCODER_NAME = "text_encoder"
+UNET_NAME = "unet"
+TRANSFORMER_NAME = "transformer"
+
+TORCH_LORA_WEIGHT_NAME = "pytorch_lora_weights.bin"
+TORCH_LORA_WEIGHT_NAME_SAFE = "pytorch_lora_weights.safetensors"
+
+PADDLE_LORA_WEIGHT_NAME = "paddle_lora_weights.pdparams"
+PADDLE_LORA_WEIGHT_NAME_SAFE = "paddle_lora_weights.safetensors"
+
+
+LORA_DEPRECATION_MESSAGE = "You are using an old version of LoRA backend. This will be deprecated in the next releases in favor of PEFT make sure to install the latest PEFT and transformers packages in the future."
+
+
+class LoraLoaderMixin:
+    r"""
+    Load LoRA layers into [`UNet2DConditionModel`] and
+    [`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel).
+    """
+
+    text_encoder_name = TEXT_ENCODER_NAME
+    unet_name = UNET_NAME
+    num_fused_loras = 0
+
+    def load_lora_weights(
+        self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, paddle.Tensor]], adapter_name=None, **kwargs
+    ):
+        """
+        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.unet` and
+        `self.text_encoder`.
+
+        All kwargs are forwarded to `self.lora_state_dict`.
+
+        See [`~loaders.LoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
+
+        See [`~loaders.LoraLoaderMixin.load_lora_into_unet`] for more details on how the state dict is loaded into
+        `self.unet`.
+
+        See [`~loaders.LoraLoaderMixin.load_lora_into_text_encoder`] for more details on how the state dict is loaded
+        into `self.text_encoder`.
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                See [`~loaders.LoraLoaderMixin.lora_state_dict`].
+            kwargs (`dict`, *optional*):
+                See [`~loaders.LoraLoaderMixin.lora_state_dict`].
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+        """
+        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
+        state_dict, network_alphas, from_diffusers = self.lora_state_dict(
+            pretrained_model_name_or_path_or_dict, **kwargs
+        )
+
+        is_correct_format = all("lora" in key for key in state_dict.keys())
+        if not is_correct_format:
+            raise ValueError("Invalid LoRA checkpoint.")
+
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", LOW_CPU_MEM_USAGE_DEFAULT)
+
+        self.load_lora_into_unet(
+            state_dict,
+            network_alphas=network_alphas,
+            unet=getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+            adapter_name=adapter_name,
+            _pipeline=self,
+            from_diffusers=from_diffusers,
+        )
+        self.load_lora_into_text_encoder(
+            state_dict,
+            network_alphas=network_alphas,
+            text_encoder=getattr(self, self.text_encoder_name)
+            if not hasattr(self, "text_encoder")
+            else self.text_encoder,
+            lora_scale=self.lora_scale,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+            adapter_name=adapter_name,
+            _pipeline=self,
+            from_diffusers=from_diffusers,
+        )
+
+    @classmethod
+    def lora_state_dict(
+        cls,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, paddle.Tensor]],
+        **kwargs,
+    ):
+        r"""
+        Return state dict for lora weights and the network alphas.
+
+        <Tip warning={true}>
+
+        We support loading A1111 formatted LoRA checkpoints in a limited capacity.
+
+        This function is experimental and might change in the future.
+
+        </Tip>
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                Can be either:
+
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            use_auth_token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+                argument to `True` will raise an error.
+            mirror (`str`, *optional*):
+                Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
+                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+                information.
+
+        """
+        # Load the main state dict first which has the LoRA layers for either of
+        # UNet and text encoder or both.
+        from_hf_hub = kwargs.pop("from_hf_hub", FROM_HF_HUB)
+        from_aistudio = kwargs.pop("from_aistudio", FROM_AISTUDIO)
+        cache_dir = kwargs.pop("cache_dir", None)
+        if cache_dir is None:
+            if from_aistudio:
+                cache_dir = None  # TODO, check aistudio cache
+            elif from_hf_hub:
+                cache_dir = DIFFUSERS_CACHE
+            else:
+                cache_dir = PPDIFFUSERS_CACHE
+        from_diffusers = kwargs.pop("from_diffusers", FROM_DIFFUSERS)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
+        revision = kwargs.pop("revision", None)
+        subfolder = kwargs.pop("subfolder", None)
+        weight_name = kwargs.pop("weight_name", None)
+        unet_config = kwargs.pop("unet_config", None)
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        if use_safetensors is None:
+            use_safetensors = True
+
+        user_agent = {
+            "file_type": "attn_procs_weights",
+            "framework": "pytorch" if from_diffusers else "paddle",
+        }
+
+        model_file = None
+        state_dict = {}
+        if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+            # Let's first try to load .safetensors weights
+            if (use_safetensors and weight_name is None) or (
+                weight_name is not None and weight_name.endswith(".safetensors")
+            ):
+                # Here we're relaxing the loading check to enable more Inference API
+                # friendliness where sometimes, it's not at all possible to automatically
+                # determine `weight_name`.
+                # if weight_name is None:
+                #     weight_name = cls._best_guess_weight_name(
+                #         pretrained_model_name_or_path_or_dict, file_extension=".safetensors"
+                #     )
+                try:
+                    model_file = _get_model_file(
+                        pretrained_model_name_or_path_or_dict,
+                        weights_name=(weight_name or TORCH_LORA_WEIGHT_NAME_SAFE)
+                        if from_diffusers
+                        else ((weight_name or PADDLE_LORA_WEIGHT_NAME_SAFE)),
+                        cache_dir=cache_dir,
+                        force_download=force_download,
+                        resume_download=resume_download,
+                        proxies=proxies,
+                        local_files_only=local_files_only,
+                        use_auth_token=use_auth_token,
+                        revision=revision,
+                        subfolder=subfolder,
+                        user_agent=user_agent,
+                        from_aistudio=from_aistudio,
+                        from_hf_hub=from_hf_hub,
+                    )
+                except Exception:
+                    model_file = None
+
+            if model_file is None:
+                # if weight_name is None:
+                #     weight_name = cls._best_guess_weight_name(
+                #         pretrained_model_name_or_path_or_dict, file_extension=".bin"
+                #     )
+                if model_file is None:
+                    model_file = _get_model_file(
+                        pretrained_model_name_or_path_or_dict,
+                        weights_name=(weight_name or TORCH_LORA_WEIGHT_NAME)
+                        if from_diffusers
+                        else ((weight_name or PADDLE_LORA_WEIGHT_NAME)),
+                        cache_dir=cache_dir,
+                        force_download=force_download,
+                        resume_download=resume_download,
+                        proxies=proxies,
+                        local_files_only=local_files_only,
+                        use_auth_token=use_auth_token,
+                        revision=revision,
+                        subfolder=subfolder,
+                        user_agent=user_agent,
+                        from_aistudio=from_aistudio,
+                        from_hf_hub=from_hf_hub,
+                    )
+
+            assert model_file is not None, "Could not find the model file!"
+            # TODO, check this
+            from ppdiffusers.utils import smart_load
+
+            state_dict = smart_load(model_file, return_is_torch_weight=True)
+            is_torch_weight = state_dict.pop("is_torch_weight", False)
+            if not from_diffusers and is_torch_weight:
+                logger.warning(
+                    "Detect the weight is in diffusers format, but currently, `from_diffusers` is set to `False`. To proceed, we will change the value of `from_diffusers` to `True`!"
+                )
+                from_diffusers = True
+        else:
+            state_dict = pretrained_model_name_or_path_or_dict
+
+        network_alphas = None
+        # TODO: replace it with a method from `state_dict_utils`
+        if all(
+            (
+                k.startswith("lora_te_")
+                or k.startswith("lora_unet_")
+                or k.startswith("lora_te1_")
+                or k.startswith("lora_te2_")
+            )
+            for k in state_dict.keys()
+        ):
+            # Map SDXL blocks correctly.
+            if unet_config is not None:
+                # use unet config to remap block numbers
+                state_dict = _maybe_map_sgm_blocks_to_diffusers(state_dict, unet_config)
+            state_dict, network_alphas = _convert_kohya_lora_to_diffusers(state_dict)
+
+        return state_dict, network_alphas, from_diffusers
+
+    @classmethod
+    def _best_guess_weight_name(cls, pretrained_model_name_or_path_or_dict, file_extension=".safetensors"):
+        targeted_files = []
+
+        if os.path.isfile(pretrained_model_name_or_path_or_dict):
+            return
+        elif os.path.isdir(pretrained_model_name_or_path_or_dict):
+            targeted_files = [
+                f for f in os.listdir(pretrained_model_name_or_path_or_dict) if f.endswith(file_extension)
+            ]
+        else:
+            files_in_repo = model_info(pretrained_model_name_or_path_or_dict).siblings
+            targeted_files = [f.rfilename for f in files_in_repo if f.rfilename.endswith(file_extension)]
+        if len(targeted_files) == 0:
+            return
+
+        # "scheduler" does not correspond to a LoRA checkpoint.
+        # "optimizer" does not correspond to a LoRA checkpoint
+        # only top-level checkpoints are considered and not the other ones, hence "checkpoint".
+        unallowed_substrings = {"scheduler", "optimizer", "checkpoint"}
+        targeted_files = list(
+            filter(lambda x: all(substring not in x for substring in unallowed_substrings), targeted_files)
+        )
+
+        if any(f.endswith(TORCH_LORA_WEIGHT_NAME) for f in targeted_files):
+            targeted_files = list(filter(lambda x: x.endswith(TORCH_LORA_WEIGHT_NAME), targeted_files))
+        elif any(f.endswith(TORCH_LORA_WEIGHT_NAME_SAFE) for f in targeted_files):
+            targeted_files = list(filter(lambda x: x.endswith(TORCH_LORA_WEIGHT_NAME_SAFE), targeted_files))
+
+        if len(targeted_files) > 1:
+            raise ValueError(
+                f"Provided path contains more than one weights file in the {file_extension} format. Either specify `weight_name` in `load_lora_weights` or make sure there's only one  `.safetensors` or `.bin` file in  {pretrained_model_name_or_path_or_dict}."
+            )
+        weight_name = targeted_files[0]
+        return weight_name
+
+    @classmethod
+    def _optionally_disable_offloading(cls, _pipeline):
+        """
+        Optionally removes offloading in case the pipeline has been already sequentially offloaded to CPU.
+
+        Args:
+            _pipeline (`DiffusionPipeline`):
+                The pipeline to disable offloading for.
+
+        Returns:
+            tuple:
+                A tuple indicating if `is_model_cpu_offload` or `is_sequential_cpu_offload` is True.
+        """
+        pass
+
+    @classmethod
+    def load_lora_into_unet(
+        cls,
+        state_dict,
+        network_alphas,
+        unet,
+        low_cpu_mem_usage=None,
+        adapter_name=None,
+        _pipeline=None,
+        from_diffusers=None,
+    ):
+        """
+        This will load the LoRA layers specified in `state_dict` into `unet`.
+
+        Parameters:
+            state_dict (`dict`):
+                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
+                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
+                encoder lora layers.
+            network_alphas (`Dict[str, float]`):
+                See `LoRALinearLayer` for more details.
+            unet (`UNet2DConditionModel`):
+                The UNet model to load the LoRA layers into.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+                argument to `True` will raise an error.
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+        """
+        if from_diffusers is None:
+            from_diffusers = FROM_DIFFUSERS
+        low_cpu_mem_usage = low_cpu_mem_usage if low_cpu_mem_usage is not None else LOW_CPU_MEM_USAGE_DEFAULT
+        # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
+        # then the `state_dict` keys should have `cls.unet_name` and/or `cls.text_encoder_name` as
+        # their prefixes.
+        keys = list(state_dict.keys())
+
+        if all(key.startswith(cls.unet_name) or key.startswith(cls.text_encoder_name) for key in keys):
+            # Load the layers corresponding to UNet.
+            logger.info(f"Loading {cls.unet_name}.")
+
+            unet_keys = [k for k in keys if k.startswith(cls.unet_name)]
+            state_dict = {k.replace(f"{cls.unet_name}.", ""): v for k, v in state_dict.items() if k in unet_keys}
+
+            if network_alphas is not None:
+                alpha_keys = [k for k in network_alphas.keys() if k.startswith(cls.unet_name)]
+                network_alphas = {
+                    k.replace(f"{cls.unet_name}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
+                }
+
+        else:
+            # Otherwise, we're dealing with the old format. This means the `state_dict` should only
+            # contain the module names of the `unet` as its keys WITHOUT any prefix.
+            warn_message = "You have saved the LoRA weights using the old format. To convert the old LoRA weights to the new format, you can first load them in a dictionary and then create a new dictionary like the following: `new_state_dict = {f'unet.{module_name}': params for module_name, params in old_state_dict.items()}`."
+            logger.warn(warn_message)
+
+        if USE_PEFT_BACKEND and len(state_dict.keys()) > 0:
+            from ppdiffusers.peft import (
+                LoraConfig,
+                inject_adapter_in_model,
+                set_peft_model_state_dict,
+            )
+
+            if adapter_name in getattr(unet, "peft_config", {}):
+                raise ValueError(
+                    f"Adapter name {adapter_name} already in use in the Unet - please select a new adapter name."
+                )
+
+            state_dict = convert_unet_state_dict_to_peft(state_dict)
+
+            if network_alphas is not None:
+                # The alphas state dict have the same structure as Unet, thus we convert it to peft format using
+                # `convert_unet_state_dict_to_peft` method.
+                network_alphas = convert_unet_state_dict_to_peft(network_alphas)
+
+            rank = {}
+            index = 1 if from_diffusers else 0
+            for key, val in state_dict.items():
+                if "lora_B" in key:
+                    rank[key] = val.shape[index]
+
+            lora_config_kwargs = get_peft_kwargs(rank, network_alphas, state_dict, is_unet=True)
+            lora_config = LoraConfig(**lora_config_kwargs)
+
+            # adapter_name
+            if adapter_name is None:
+                adapter_name = get_adapter_name(unet)
+
+            # In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
+            # otherwise loading LoRA weights will lead to an error
+
+            inject_adapter_in_model(lora_config, unet, adapter_name=adapter_name)
+            incompatible_keys = set_peft_model_state_dict(
+                unet, state_dict, adapter_name, from_diffusers=from_diffusers
+            )  # NOTE: new add from_diffusers
+
+            if incompatible_keys is not None:
+                # check only for unexpected keys
+                unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
+                if unexpected_keys:
+                    logger.warning(
+                        f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
+                        f" {unexpected_keys}. "
+                    )
+
+        unet.load_attn_procs(
+            state_dict,
+            network_alphas=network_alphas,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+            _pipeline=_pipeline,
+            from_diffusers=from_diffusers,
+        )
+
+    @classmethod
+    def load_lora_into_text_encoder(
+        cls,
+        state_dict,
+        network_alphas,
+        text_encoder,
+        prefix=None,
+        lora_scale=1.0,
+        low_cpu_mem_usage=None,
+        adapter_name=None,
+        _pipeline=None,
+        from_diffusers=None,
+    ):
+        """
+        This will load the LoRA layers specified in `state_dict` into `text_encoder`
+
+        Parameters:
+            state_dict (`dict`):
+                A standard state dict containing the lora layer parameters. The key should be prefixed with an
+                additional `text_encoder` to distinguish between unet lora layers.
+            network_alphas (`Dict[str, float]`):
+                See `LoRALinearLayer` for more details.
+            text_encoder (`CLIPTextModel`):
+                The text encoder model to load the LoRA layers into.
+            prefix (`str`):
+                Expected prefix of the `text_encoder` in the `state_dict`.
+            lora_scale (`float`):
+                How much to scale the output of the lora linear layer before it is added with the output of the regular
+                lora layer.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+                argument to `True` will raise an error.
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+        """
+        if from_diffusers is None:
+            from_diffusers = FROM_DIFFUSERS
+        low_cpu_mem_usage = low_cpu_mem_usage if low_cpu_mem_usage is not None else LOW_CPU_MEM_USAGE_DEFAULT
+
+        # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
+        # then the `state_dict` keys should have `self.unet_name` and/or `self.text_encoder_name` as
+        # their prefixes.
+        keys = list(state_dict.keys())
+        prefix = cls.text_encoder_name if prefix is None else prefix
+
+        # Safe prefix to check with.
+        if any(cls.text_encoder_name in key for key in keys):
+            # Load the layers corresponding to text encoder and make necessary adjustments.
+            text_encoder_keys = [k for k in keys if k.startswith(prefix) and k.split(".")[0] == prefix]
+            text_encoder_lora_state_dict = {
+                k.replace(f"{prefix}.", ""): v for k, v in state_dict.items() if k in text_encoder_keys
+            }
+
+            if len(text_encoder_lora_state_dict) > 0:
+                logger.info(f"Loading {prefix}.")
+                rank = {}
+                text_encoder_lora_state_dict = convert_state_dict_to_ppdiffusers(text_encoder_lora_state_dict)
+                index = 1 if from_diffusers else 0
+                if USE_PEFT_BACKEND:
+                    # convert state dict
+                    text_encoder_lora_state_dict = convert_state_dict_to_peft(text_encoder_lora_state_dict)
+
+                    for name, _ in text_encoder_attn_modules(text_encoder):
+                        rank_key = f"{name}.out_proj.lora_B.weight"
+                        rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[index]
+
+                    patch_mlp = any(".mlp." in key for key in text_encoder_lora_state_dict.keys())
+                    if patch_mlp:
+                        for name, _ in text_encoder_mlp_modules(text_encoder):
+                            rank_key_fc1 = f"{name}.fc1.lora_B.weight"
+                            rank_key_fc2 = f"{name}.fc2.lora_B.weight"
+
+                            rank[rank_key_fc1] = text_encoder_lora_state_dict[rank_key_fc1].shape[index]
+                            rank[rank_key_fc2] = text_encoder_lora_state_dict[rank_key_fc2].shape[index]
+                else:
+                    for name, _ in text_encoder_attn_modules(text_encoder):
+                        rank_key = f"{name}.out_proj.lora_linear_layer.up.weight"
+                        rank.update({rank_key: text_encoder_lora_state_dict[rank_key].shape[index]})
+
+                    patch_mlp = any(".mlp." in key for key in text_encoder_lora_state_dict.keys())
+                    if patch_mlp:
+                        for name, _ in text_encoder_mlp_modules(text_encoder):
+                            rank_key_fc1 = f"{name}.fc1.lora_linear_layer.up.weight"
+                            rank_key_fc2 = f"{name}.fc2.lora_linear_layer.up.weight"
+                            rank[rank_key_fc1] = text_encoder_lora_state_dict[rank_key_fc1].shape[index]
+                            rank[rank_key_fc2] = text_encoder_lora_state_dict[rank_key_fc2].shape[index]
+                if network_alphas is not None:
+                    alpha_keys = [
+                        k for k in network_alphas.keys() if k.startswith(prefix) and k.split(".")[0] == prefix
+                    ]
+                    network_alphas = {
+                        k.replace(f"{prefix}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
+                    }
+
+                if USE_PEFT_BACKEND:
+                    from ppdiffusers.peft import LoraConfig
+
+                    lora_config_kwargs = get_peft_kwargs(
+                        rank, network_alphas, text_encoder_lora_state_dict, is_unet=False
+                    )
+
+                    lora_config = LoraConfig(**lora_config_kwargs)
+
+                    # adapter_name
+                    if adapter_name is None:
+                        adapter_name = get_adapter_name(text_encoder)
+
+                    # inject LoRA layers and load the state dict
+                    # in transformers we automatically check whether the adapter name is already in use or not
+                    text_encoder.load_adapter(
+                        adapter_name=adapter_name,
+                        adapter_state_dict=text_encoder_lora_state_dict,
+                        peft_config=lora_config,
+                        from_diffusers=from_diffusers,  # new add from diffusers
+                    )
+
+                    # scale LoRA layers with `lora_scale`
+                    scale_lora_layers(text_encoder, weight=lora_scale)
+                else:
+                    cls._modify_text_encoder(
+                        text_encoder,
+                        lora_scale,
+                        network_alphas,
+                        rank=rank,
+                        patch_mlp=patch_mlp,
+                        low_cpu_mem_usage=low_cpu_mem_usage,
+                    )
+
+                    if from_diffusers:
+                        convert_pytorch_state_dict_to_paddle(text_encoder, text_encoder_lora_state_dict)
+                    faster_set_state_dict(text_encoder, text_encoder_lora_state_dict)
+
+                text_encoder.to(dtype=text_encoder.dtype)
+
+    @property
+    def lora_scale(self) -> float:
+        # property function that returns the lora scale which can be set at run time by the pipeline.
+        # if _lora_scale has not been set, return 1
+        return self._lora_scale if hasattr(self, "_lora_scale") else 1.0
+
+    def _remove_text_encoder_monkey_patch(self):
+        if USE_PEFT_BACKEND:
+            remove_method = recurse_remove_peft_layers
+        else:
+            remove_method = self._remove_text_encoder_monkey_patch_classmethod
+
+        if hasattr(self, "text_encoder"):
+            remove_method(self.text_encoder)
+
+            # In case text encoder have no Lora attached
+            if USE_PEFT_BACKEND and getattr(self.text_encoder, "peft_config", None) is not None:
+                del self.text_encoder.peft_config
+                self.text_encoder._pp_peft_config_loaded = None
+        if hasattr(self, "text_encoder_2"):
+            remove_method(self.text_encoder_2)
+            if USE_PEFT_BACKEND:
+                del self.text_encoder_2.peft_config
+                self.text_encoder_2._pp_peft_config_loaded = None
+
+    @classmethod
+    def _remove_text_encoder_monkey_patch_classmethod(cls, text_encoder):
+        if version.parse(__version__) > version.parse("0.23"):
+            deprecate("_remove_text_encoder_monkey_patch_classmethod", "0.25", LORA_DEPRECATION_MESSAGE)
+
+        for _, attn_module in text_encoder_attn_modules(text_encoder):
+            if isinstance(attn_module.q_proj, PatchedLoraProjection):
+                attn_module.q_proj.lora_linear_layer = None
+                attn_module.k_proj.lora_linear_layer = None
+                attn_module.v_proj.lora_linear_layer = None
+                attn_module.out_proj.lora_linear_layer = None
+
+        for _, mlp_module in text_encoder_mlp_modules(text_encoder):
+            if isinstance(mlp_module.fc1, PatchedLoraProjection):
+                mlp_module.fc1.lora_linear_layer = None
+                mlp_module.fc2.lora_linear_layer = None
+
+    @classmethod
+    def _modify_text_encoder(
+        cls,
+        text_encoder,
+        lora_scale=1,
+        network_alphas=None,
+        rank: Union[Dict[str, int], int] = 4,
+        dtype=None,
+        patch_mlp=False,
+        low_cpu_mem_usage=False,
+    ):
+        r"""
+        Monkey-patches the forward passes of attention modules of the text encoder.
+        """
+        if version.parse(__version__) > version.parse("0.23"):
+            deprecate("_modify_text_encoder", "0.25", LORA_DEPRECATION_MESSAGE)
+
+        def create_patched_linear_lora(model, network_alpha, rank, dtype, lora_parameters):
+            linear_layer = model.regular_linear_layer if isinstance(model, PatchedLoraProjection) else model
+            ctx = paddle.LazyGuard if low_cpu_mem_usage else nullcontext
+            with ctx():
+                model = PatchedLoraProjection(linear_layer, lora_scale, network_alpha, rank, dtype=dtype)
+
+            lora_parameters.extend(model.lora_linear_layer.parameters())
+            return model
+
+        # First, remove any monkey-patch that might have been applied before
+        cls._remove_text_encoder_monkey_patch_classmethod(text_encoder)
+
+        lora_parameters = []
+        network_alphas = {} if network_alphas is None else network_alphas
+        is_network_alphas_populated = len(network_alphas) > 0
+
+        for name, attn_module in text_encoder_attn_modules(text_encoder):
+            query_alpha = network_alphas.pop(name + ".to_q_lora.down.weight.alpha", None)
+            key_alpha = network_alphas.pop(name + ".to_k_lora.down.weight.alpha", None)
+            value_alpha = network_alphas.pop(name + ".to_v_lora.down.weight.alpha", None)
+            out_alpha = network_alphas.pop(name + ".to_out_lora.down.weight.alpha", None)
+
+            if isinstance(rank, dict):
+                current_rank = rank.pop(f"{name}.out_proj.lora_linear_layer.up.weight")
+            else:
+                current_rank = rank
+
+            attn_module.q_proj = create_patched_linear_lora(
+                attn_module.q_proj, query_alpha, current_rank, dtype, lora_parameters
+            )
+            attn_module.k_proj = create_patched_linear_lora(
+                attn_module.k_proj, key_alpha, current_rank, dtype, lora_parameters
+            )
+            attn_module.v_proj = create_patched_linear_lora(
+                attn_module.v_proj, value_alpha, current_rank, dtype, lora_parameters
+            )
+            attn_module.out_proj = create_patched_linear_lora(
+                attn_module.out_proj, out_alpha, current_rank, dtype, lora_parameters
+            )
+
+        if patch_mlp:
+            for name, mlp_module in text_encoder_mlp_modules(text_encoder):
+                fc1_alpha = network_alphas.pop(name + ".fc1.lora_linear_layer.down.weight.alpha", None)
+                fc2_alpha = network_alphas.pop(name + ".fc2.lora_linear_layer.down.weight.alpha", None)
+
+                current_rank_fc1 = rank.pop(f"{name}.fc1.lora_linear_layer.up.weight")
+                current_rank_fc2 = rank.pop(f"{name}.fc2.lora_linear_layer.up.weight")
+
+                mlp_module.fc1 = create_patched_linear_lora(
+                    mlp_module.fc1, fc1_alpha, current_rank_fc1, dtype, lora_parameters
+                )
+                mlp_module.fc2 = create_patched_linear_lora(
+                    mlp_module.fc2, fc2_alpha, current_rank_fc2, dtype, lora_parameters
+                )
+
+        if is_network_alphas_populated and len(network_alphas) > 0:
+            raise ValueError(
+                f"The `network_alphas` has to be empty at this point but has the following keys \n\n {', '.join(network_alphas.keys())}"
+            )
+
+        return lora_parameters
+
+    @classmethod
+    def save_lora_weights(
+        cls,
+        save_directory: Union[str, os.PathLike],
+        unet_lora_layers: Dict[str, Union[nn.Layer, paddle.Tensor]] = None,
+        text_encoder_lora_layers: Dict[str, nn.Layer] = None,
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+        to_diffusers=None,
+    ):
+        r"""
+        Save the LoRA parameters corresponding to the UNet and text encoder.
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to save LoRA parameters to. Will be created if it doesn't exist.
+            unet_lora_layers (`Dict[str, nn.Layer]` or `Dict[str, paddle.Tensor]`):
+                State dict of the LoRA layers corresponding to the `unet`.
+            text_encoder_lora_layers (`Dict[str, nn.Layer]` or `Dict[str, paddle.Tensor]`):
+                State dict of the LoRA layers corresponding to the `text_encoder`. Must explicitly pass the text
+                encoder LoRA state dict because it comes from 🤗 Transformers.
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful during distributed training and you
+                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+                process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful during distributed training when you need to
+                replace `torch.save` with another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
+        """
+        if to_diffusers is None:
+            to_diffusers = TO_DIFFUSERS
+        # Create a flat dictionary.
+        state_dict = {}
+
+        # Populate the dictionary.
+        if unet_lora_layers is not None:
+            weights = unet_lora_layers.state_dict() if isinstance(unet_lora_layers, nn.Layer) else unet_lora_layers
+            if to_diffusers and isinstance(unet_lora_layers, nn.Layer):
+                convert_paddle_state_dict_to_pytorch(unet_lora_layers, weights)
+
+            unet_lora_state_dict = {f"{cls.unet_name}.{module_name}": param for module_name, param in weights.items()}
+
+            state_dict.update(unet_lora_state_dict)
+
+        if text_encoder_lora_layers is not None:
+            weights = (
+                text_encoder_lora_layers.state_dict()
+                if isinstance(text_encoder_lora_layers, nn.Layer)
+                else text_encoder_lora_layers
+            )
+            if to_diffusers and isinstance(text_encoder_lora_layers, nn.Layer):
+                convert_paddle_state_dict_to_pytorch(text_encoder_lora_layers, weights)
+
+            text_encoder_lora_state_dict = {
+                f"{cls.text_encoder_name}.{module_name}": param for module_name, param in weights.items()
+            }
+            state_dict.update(text_encoder_lora_state_dict)
+
+        # Save the model
+        cls.write_lora_layers(
+            state_dict=state_dict,
+            save_directory=save_directory,
+            is_main_process=is_main_process,
+            weight_name=weight_name,
+            save_function=save_function,
+            safe_serialization=safe_serialization,
+            to_diffusers=to_diffusers,  # only change save weights name and save function
+        )
+
+    @staticmethod
+    def write_lora_layers(
+        state_dict: Dict[str, paddle.Tensor],
+        save_directory: str,
+        is_main_process: bool,
+        weight_name: str,
+        save_function: Callable,
+        safe_serialization: bool,
+        to_diffusers=None,
+    ):
+        if to_diffusers is None:
+            to_diffusers = TO_DIFFUSERS
+        if os.path.isfile(save_directory):
+            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
+            return
+
+        os.makedirs(save_directory, exist_ok=True)
+
+        if weight_name is None:
+            if to_diffusers:
+                if safe_serialization:
+                    weight_name = TORCH_LORA_WEIGHT_NAME_SAFE
+                else:
+                    weight_name = TORCH_LORA_WEIGHT_NAME
+            else:
+                if safe_serialization:
+                    weight_name = PADDLE_LORA_WEIGHT_NAME_SAFE
+                else:
+                    weight_name = PADDLE_LORA_WEIGHT_NAME
+        else:
+            if "paddle" in weight_name.lower() or "pdparams" in weight_name.lower():
+                to_diffusers = False
+            elif "torch" in weight_name.lower() or "bin" in weight_name.lower():
+                to_diffusers = True
+
+        # choose save_function
+        if save_function is None:
+            if to_diffusers:
+                if not is_torch_available() and not safe_serialization:
+                    safe_serialization = True
+                    logger.warning(
+                        "PyTorch is not installed, and `safe_serialization` is currently set to `False`. "
+                        "To ensure proper model saving, we will automatically set `safe_serialization=True`. "
+                        "If you want to keep `safe_serialization=False`, please make sure PyTorch is installed."
+                    )
+                if safe_serialization:
+                    if is_torch_available():
+                        save_function = partial(torch_safe_save_file, metadata={"format": "pt"})
+                    else:
+                        save_function = partial(np_safe_save_file, metadata={"format": "pt"})
+                else:
+                    save_function = torch.save
+            else:
+                if safe_serialization:
+                    state_dict = {k: np.ascontiguousarray(v) for k, v in state_dict.items()}
+                    save_function = partial(np_safe_save_file, metadata={"format": "pd"})
+                else:
+                    save_function = paddle.save
+
+        # we have transpose state_dict!
+
+        save_function(state_dict, os.path.join(save_directory, weight_name))
+        logger.info(f"Model weights saved in {os.path.join(save_directory, weight_name)}")
+
+    def unload_lora_weights(self):
+        """
+        Unloads the LoRA parameters.
+
+        Examples:
+
+        ```python
+        >>> # Assuming `pipeline` is already loaded with the LoRA parameters.
+        >>> pipeline.unload_lora_weights()
+        >>> ...
+        ```
+        """
+        if not USE_PEFT_BACKEND:
+            if version.parse(__version__) > version.parse("0.23"):
+                logger.warn(
+                    "You are using `unload_lora_weights` to disable and unload lora weights. If you want to iteratively enable and disable adapter weights,"
+                    "you can use `pipe.enable_lora()` or `pipe.disable_lora()`. After installing the latest version of PEFT."
+                )
+
+            for _, module in self.unet.named_sublayers(include_self=True):
+                if hasattr(module, "set_lora_layer"):
+                    module.set_lora_layer(None)
+        else:
+            recurse_remove_peft_layers(self.unet)
+            if hasattr(self.unet, "peft_config"):
+                del self.unet.peft_config
+
+        # Safe to call the following regardless of LoRA.
+        self._remove_text_encoder_monkey_patch()
+
+    def fuse_lora(
+        self,
+        fuse_unet: bool = True,
+        fuse_text_encoder: bool = True,
+        lora_scale: float = 1.0,
+        safe_fusing: bool = False,
+    ):
+        r"""
+        Fuses the LoRA parameters into the original parameters of the corresponding blocks.
+
+        <Tip warning={true}>
+
+        This is an experimental API.
+
+        </Tip>
+
+        Args:
+            fuse_unet (`bool`, defaults to `True`): Whether to fuse the UNet LoRA parameters.
+            fuse_text_encoder (`bool`, defaults to `True`):
+                Whether to fuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
+                LoRA parameters then it won't have any effect.
+            lora_scale (`float`, defaults to 1.0):
+                Controls how much to influence the outputs with the LoRA parameters.
+            safe_fusing (`bool`, defaults to `False`):
+                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
+        """
+        if fuse_unet or fuse_text_encoder:
+            self.num_fused_loras += 1
+            if self.num_fused_loras > 1:
+                logger.warn(
+                    "The current API is supported for operating with a single LoRA file. You are trying to load and fuse more than one LoRA which is not well-supported.",
+                )
+
+        if fuse_unet:
+            self.unet.fuse_lora(lora_scale, safe_fusing=safe_fusing)
+
+        if USE_PEFT_BACKEND:
+            from ppdiffusers.peft.tuners.tuners_utils import BaseTunerLayer
+
+            def fuse_text_encoder_lora(text_encoder, lora_scale=1.0, safe_fusing=False):
+                # TODO(Patrick, Younes): enable "safe" fusing
+                for module in text_encoder.sublayers(include_self=True):
+                    if isinstance(module, BaseTunerLayer):
+                        if lora_scale != 1.0:
+                            module.scale_layer(lora_scale)
+
+                        module.merge()
+
+        else:
+            if version.parse(__version__) > version.parse("0.23"):
+                deprecate("fuse_text_encoder_lora", "0.25", LORA_DEPRECATION_MESSAGE)
+
+            def fuse_text_encoder_lora(text_encoder, lora_scale=1.0, safe_fusing=False):
+                for _, attn_module in text_encoder_attn_modules(text_encoder):
+                    if isinstance(attn_module.q_proj, PatchedLoraProjection):
+                        attn_module.q_proj._fuse_lora(lora_scale, safe_fusing)
+                        attn_module.k_proj._fuse_lora(lora_scale, safe_fusing)
+                        attn_module.v_proj._fuse_lora(lora_scale, safe_fusing)
+                        attn_module.out_proj._fuse_lora(lora_scale, safe_fusing)
+
+                for _, mlp_module in text_encoder_mlp_modules(text_encoder):
+                    if isinstance(mlp_module.fc1, PatchedLoraProjection):
+                        mlp_module.fc1._fuse_lora(lora_scale, safe_fusing)
+                        mlp_module.fc2._fuse_lora(lora_scale, safe_fusing)
+
+        if fuse_text_encoder:
+            if hasattr(self, "text_encoder"):
+                fuse_text_encoder_lora(self.text_encoder, lora_scale, safe_fusing)
+            if hasattr(self, "text_encoder_2"):
+                fuse_text_encoder_lora(self.text_encoder_2, lora_scale, safe_fusing)
+
+    def unfuse_lora(self, unfuse_unet: bool = True, unfuse_text_encoder: bool = True):
+        r"""
+        Reverses the effect of
+        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.fuse_lora).
+
+        <Tip warning={true}>
+
+        This is an experimental API.
+
+        </Tip>
+
+        Args:
+            unfuse_unet (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
+            unfuse_text_encoder (`bool`, defaults to `True`):
+                Whether to unfuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
+                LoRA parameters then it won't have any effect.
+        """
+        if unfuse_unet:
+            if not USE_PEFT_BACKEND:
+                self.unet.unfuse_lora()
+            else:
+                from ppdiffusers.peft.tuners.tuners_utils import BaseTunerLayer
+
+                for module in self.unet.sublayers(include_self=True):
+                    if isinstance(module, BaseTunerLayer):
+                        module.unmerge()
+
+        if USE_PEFT_BACKEND:
+            from ppdiffusers.peft.tuners.tuners_utils import BaseTunerLayer
+
+            def unfuse_text_encoder_lora(text_encoder):
+                for module in text_encoder.sublayers(include_self=True):
+                    if isinstance(module, BaseTunerLayer):
+                        module.unmerge()
+
+        else:
+            if version.parse(__version__) > version.parse("0.23"):
+                deprecate("unfuse_text_encoder_lora", "0.25", LORA_DEPRECATION_MESSAGE)
+
+            def unfuse_text_encoder_lora(text_encoder):
+                for _, attn_module in text_encoder_attn_modules(text_encoder):
+                    if isinstance(attn_module.q_proj, PatchedLoraProjection):
+                        attn_module.q_proj._unfuse_lora()
+                        attn_module.k_proj._unfuse_lora()
+                        attn_module.v_proj._unfuse_lora()
+                        attn_module.out_proj._unfuse_lora()
+
+                for _, mlp_module in text_encoder_mlp_modules(text_encoder):
+                    if isinstance(mlp_module.fc1, PatchedLoraProjection):
+                        mlp_module.fc1._unfuse_lora()
+                        mlp_module.fc2._unfuse_lora()
+
+        if unfuse_text_encoder:
+            if hasattr(self, "text_encoder"):
+                unfuse_text_encoder_lora(self.text_encoder)
+            if hasattr(self, "text_encoder_2"):
+                unfuse_text_encoder_lora(self.text_encoder_2)
+
+        self.num_fused_loras -= 1
+
+    def set_adapters_for_text_encoder(
+        self,
+        adapter_names: Union[List[str], str],
+        text_encoder: Optional["PretrainedModel"] = None,  # noqa: F821
+        text_encoder_weights: List[float] = None,
+    ):
+        """
+        Sets the adapter layers for the text encoder.
+
+        Args:
+            adapter_names (`List[str]` or `str`):
+                The names of the adapters to use.
+            text_encoder (`nn.Layer`, *optional*):
+                The text encoder module to set the adapter layers for. If `None`, it will try to get the `text_encoder`
+                attribute.
+            text_encoder_weights (`List[float]`, *optional*):
+                The weights to use for the text encoder. If `None`, the weights are set to `1.0` for all the adapters.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        def process_weights(adapter_names, weights):
+            if weights is None:
+                weights = [1.0] * len(adapter_names)
+            elif isinstance(weights, float):
+                weights = [weights]
+
+            if len(adapter_names) != len(weights):
+                raise ValueError(
+                    f"Length of adapter names {len(adapter_names)} is not equal to the length of the weights {len(weights)}"
+                )
+            return weights
+
+        adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
+        text_encoder_weights = process_weights(adapter_names, text_encoder_weights)
+        text_encoder = text_encoder or getattr(self, "text_encoder", None)
+        if text_encoder is None:
+            raise ValueError(
+                "The pipeline does not have a default `pipe.text_encoder` class. Please make sure to pass a `text_encoder` instead."
+            )
+        set_weights_and_activate_adapters(text_encoder, adapter_names, text_encoder_weights)
+
+    def disable_lora_for_text_encoder(self, text_encoder: Optional["PretrainedModel"] = None):
+        """
+        Disables the LoRA layers for the text encoder.
+
+        Args:
+            text_encoder (`nn.Layer`, *optional*):
+                The text encoder module to disable the LoRA layers for. If `None`, it will try to get the
+                `text_encoder` attribute.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        text_encoder = text_encoder or getattr(self, "text_encoder", None)
+        if text_encoder is None:
+            raise ValueError("Text Encoder not found.")
+        set_adapter_layers(text_encoder, enabled=False)
+
+    def enable_lora_for_text_encoder(self, text_encoder: Optional["PretrainedModel"] = None):
+        """
+        Enables the LoRA layers for the text encoder.
+
+        Args:
+            text_encoder (`nn.Layer`, *optional*):
+                The text encoder module to enable the LoRA layers for. If `None`, it will try to get the `text_encoder`
+                attribute.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+        text_encoder = text_encoder or getattr(self, "text_encoder", None)
+        if text_encoder is None:
+            raise ValueError("Text Encoder not found.")
+        set_adapter_layers(self.text_encoder, enabled=True)
+
+    def set_adapters(
+        self,
+        adapter_names: Union[List[str], str],
+        adapter_weights: Optional[List[float]] = None,
+    ):
+        # Handle the UNET
+        self.unet.set_adapters(adapter_names, adapter_weights)
+
+        # Handle the Text Encoder
+        if hasattr(self, "text_encoder"):
+            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder, adapter_weights)
+        if hasattr(self, "text_encoder_2"):
+            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder_2, adapter_weights)
+
+    def disable_lora(self):
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        # Disable unet adapters
+        self.unet.disable_lora()
+
+        # Disable text encoder adapters
+        if hasattr(self, "text_encoder"):
+            self.disable_lora_for_text_encoder(self.text_encoder)
+        if hasattr(self, "text_encoder_2"):
+            self.disable_lora_for_text_encoder(self.text_encoder_2)
+
+    def enable_lora(self):
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        # Enable unet adapters
+        self.unet.enable_lora()
+
+        # Enable text encoder adapters
+        if hasattr(self, "text_encoder"):
+            self.enable_lora_for_text_encoder(self.text_encoder)
+        if hasattr(self, "text_encoder_2"):
+            self.enable_lora_for_text_encoder(self.text_encoder_2)
+
+    def delete_adapters(self, adapter_names: Union[List[str], str]):
+        """
+        Args:
+        Deletes the LoRA layers of `adapter_name` for the unet and text-encoder(s).
+            adapter_names (`Union[List[str], str]`):
+                The names of the adapter to delete. Can be a single string or a list of strings
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        if isinstance(adapter_names, str):
+            adapter_names = [adapter_names]
+
+        # Delete unet adapters
+        self.unet.delete_adapters(adapter_names)
+
+        for adapter_name in adapter_names:
+            # Delete text encoder adapters
+            if hasattr(self, "text_encoder"):
+                delete_adapter_layers(self.text_encoder, adapter_name)
+            if hasattr(self, "text_encoder_2"):
+                delete_adapter_layers(self.text_encoder_2, adapter_name)
+
+    def get_active_adapters(self) -> List[str]:
+        """
+        Gets the list of the current active adapters.
+
+        Example:
+
+        ```python
+        from ppdiffusers import DiffusionPipeline
+
+        pipeline = DiffusionPipeline.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0",
+        )
+        pipeline.load_lora_weights("CiroN2022/toy-face", weight_name="toy_face_sdxl.safetensors", adapter_name="toy")
+        pipeline.get_active_adapters()
+        ```
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError(
+                "PEFT backend is required for this method. Please install the latest version of PEFT `pip install -U peft`"
+            )
+
+        from ppdiffusers.peft.tuners.tuners_utils import BaseTunerLayer
+
+        active_adapters = []
+
+        for module in self.unet.sublayers(include_self=True):
+            if isinstance(module, BaseTunerLayer):
+                active_adapters = module.active_adapters
+                break
+
+        return active_adapters
+
+    def get_list_adapters(self) -> Dict[str, List[str]]:
+        """
+        Gets the current list of all available adapters in the pipeline.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError(
+                "PEFT backend is required for this method. Please install the latest version of PEFT `pip install -U peft`"
+            )
+
+        set_adapters = {}
+
+        if hasattr(self, "text_encoder") and hasattr(self.text_encoder, "peft_config"):
+            set_adapters["text_encoder"] = list(self.text_encoder.peft_config.keys())
+
+        if hasattr(self, "text_encoder_2") and hasattr(self.text_encoder_2, "peft_config"):
+            set_adapters["text_encoder_2"] = list(self.text_encoder_2.peft_config.keys())
+
+        if hasattr(self, "unet") and hasattr(self.unet, "peft_config"):
+            set_adapters["unet"] = list(self.unet.peft_config.keys())
+
+        return set_adapters
+
+    def set_lora_device(self, adapter_names: List[str], device: Union[str, int]) -> None:
+        """
+        Moves the LoRAs listed in `adapter_names` to a target device. Useful for offloading the LoRA to the CPU in case
+        you want to load multiple adapters and free some GPU memory.
+
+        Args:
+            adapter_names (`List[str]`):
+                List of adapters to send device to.
+            device (`Union[str, int]`):
+                Device to send the adapters to. Can be either a torch device, a str or an integer.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        # from ppdiffusers.peft.tuners.tuners_utils import BaseTunerLayer
+
+        # Handle the UNET
+        # for unet_module in self.unet.sublayers(include_self=True):
+        #     if isinstance(unet_module, BaseTunerLayer):
+        #         for adapter_name in adapter_names:
+        #             unet_module.lora_A[adapter_name].to(device)
+        #             unet_module.lora_B[adapter_name].to(device)
+
+        # Handle the text encoder
+        # modules_to_process = []
+        # if hasattr(self, "text_encoder"):
+        #     modules_to_process.append(self.text_encoder)
+
+        # if hasattr(self, "text_encoder_2"):
+        #     modules_to_process.append(self.text_encoder_2)
+
+        # for text_encoder in modules_to_process:
+        #     # loop over submodules
+        #     for text_encoder_module in text_encoder.sublayers(include_self=True):
+        #         if isinstance(text_encoder_module, BaseTunerLayer):
+        #             for adapter_name in adapter_names:
+        #                 text_encoder_module.lora_A[adapter_name].to(device)
+        #                 text_encoder_module.lora_B[adapter_name].to(device)
+
+
+class StableDiffusionXLLoraLoaderMixin(LoraLoaderMixin):
+    """This class overrides `LoraLoaderMixin` with LoRA loading/saving code that's specific to SDXL"""
+
+    # Overrride to properly handle the loading and unloading of the additional text encoder.
+    def load_lora_weights(
+        self,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, paddle.Tensor]],
+        adapter_name: Optional[str] = None,
+        **kwargs,
+    ):
+        """
+        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.unet` and
+        `self.text_encoder`.
+
+        All kwargs are forwarded to `self.lora_state_dict`.
+
+        See [`~loaders.LoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
+
+        See [`~loaders.LoraLoaderMixin.load_lora_into_unet`] for more details on how the state dict is loaded into
+        `self.unet`.
+
+        See [`~loaders.LoraLoaderMixin.load_lora_into_text_encoder`] for more details on how the state dict is loaded
+        into `self.text_encoder`.
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                See [`~loaders.LoraLoaderMixin.lora_state_dict`].
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+            kwargs (`dict`, *optional*):
+                See [`~loaders.LoraLoaderMixin.lora_state_dict`].
+        """
+        # We could have accessed the unet config from `lora_state_dict()` too. We pass
+        # it here explicitly to be able to tell that it's coming from an SDXL
+        # pipeline.
+
+        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
+        state_dict, network_alphas, from_diffusers = self.lora_state_dict(
+            pretrained_model_name_or_path_or_dict,
+            unet_config=self.unet.config,
+            **kwargs,
+        )
+        is_correct_format = all("lora" in key for key in state_dict.keys())
+        if not is_correct_format:
+            raise ValueError("Invalid LoRA checkpoint.")
+
+        self.load_lora_into_unet(
+            state_dict,
+            network_alphas=network_alphas,
+            unet=self.unet,
+            adapter_name=adapter_name,
+            _pipeline=self,
+            from_diffusers=from_diffusers,
+        )
+        text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
+        if len(text_encoder_state_dict) > 0:
+            self.load_lora_into_text_encoder(
+                text_encoder_state_dict,
+                network_alphas=network_alphas,
+                text_encoder=self.text_encoder,
+                prefix="text_encoder",
+                lora_scale=self.lora_scale,
+                adapter_name=adapter_name,
+                _pipeline=self,
+                from_diffusers=from_diffusers,
+            )
+
+        text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
+        if len(text_encoder_2_state_dict) > 0:
+            self.load_lora_into_text_encoder(
+                text_encoder_2_state_dict,
+                network_alphas=network_alphas,
+                text_encoder=self.text_encoder_2,
+                prefix="text_encoder_2",
+                lora_scale=self.lora_scale,
+                adapter_name=adapter_name,
+                _pipeline=self,
+                from_diffusers=from_diffusers,
+            )
+
+    @classmethod
+    def save_lora_weights(
+        cls,
+        save_directory: Union[str, os.PathLike],
+        unet_lora_layers: Dict[str, Union[nn.Layer, paddle.Tensor]] = None,
+        text_encoder_lora_layers: Dict[str, Union[nn.Layer, paddle.Tensor]] = None,
+        text_encoder_2_lora_layers: Dict[str, Union[nn.Layer, paddle.Tensor]] = None,
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+        to_diffusers: bool = None,
+    ):
+        r"""
+        Save the LoRA parameters corresponding to the UNet and text encoder.
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to save LoRA parameters to. Will be created if it doesn't exist.
+            unet_lora_layers (`Dict[str, nn.Layer]` or `Dict[str, paddle.Tensor]`):
+                State dict of the LoRA layers corresponding to the `unet`.
+            text_encoder_lora_layers (`Dict[str, nn.Layer]` or `Dict[str, paddle.Tensor]`):
+                State dict of the LoRA layers corresponding to the `text_encoder`. Must explicitly pass the text
+                encoder LoRA state dict because it comes from 🤗 Transformers.
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful during distributed training and you
+                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+                process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful during distributed training when you need to
+                replace `torch.save` with another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
+        """
+        if to_diffusers is None:
+            to_diffusers = TO_DIFFUSERS
+
+        state_dict = {}
+
+        def pack_weights(layers, prefix):
+            layers_weights = layers.state_dict() if isinstance(layers, nn.Layer) else layers
+            if to_diffusers and isinstance(layers, nn.Layer):
+                convert_paddle_state_dict_to_pytorch(layers, layers_weights)
+            layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
+            return layers_state_dict
+
+        if not (unet_lora_layers or text_encoder_lora_layers or text_encoder_2_lora_layers):
+            raise ValueError(
+                "You must pass at least one of `unet_lora_layers`, `text_encoder_lora_layers` or `text_encoder_2_lora_layers`."
+            )
+
+        if unet_lora_layers:
+            state_dict.update(pack_weights(unet_lora_layers, "unet"))
+
+        if text_encoder_lora_layers and text_encoder_2_lora_layers:
+            state_dict.update(pack_weights(text_encoder_lora_layers, "text_encoder"))
+            state_dict.update(pack_weights(text_encoder_2_lora_layers, "text_encoder_2"))
+
+        cls.write_lora_layers(
+            state_dict=state_dict,
+            save_directory=save_directory,
+            is_main_process=is_main_process,
+            weight_name=weight_name,
+            save_function=save_function,
+            safe_serialization=safe_serialization,
+            to_diffusers=to_diffusers,
+        )
+
+    def _remove_text_encoder_monkey_patch(self):
+        if USE_PEFT_BACKEND:
+            recurse_remove_peft_layers(self.text_encoder)
+            # TODO: @younesbelkada handle this in transformers side
+            if getattr(self.text_encoder, "peft_config", None) is not None:
+                del self.text_encoder.peft_config
+                self.text_encoder._pp_peft_config_loaded = None
+
+            recurse_remove_peft_layers(self.text_encoder_2)
+            if getattr(self.text_encoder_2, "peft_config", None) is not None:
+                del self.text_encoder_2.peft_config
+                self.text_encoder_2._pp_peft_config_loaded = None
+        else:
+            self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder)
+            self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder_2)
+
+
+class SD3LoraLoaderMixin:
+    r"""
+    Load LoRA layers into [`SD3Transformer2DModel`].
+    """
+
+    transformer_name = TRANSFORMER_NAME
+    num_fused_loras = 0
+
+    def load_lora_weights(
+        self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, paddle.Tensor]], adapter_name=None, **kwargs
+    ):
+        """
+        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.unet` and
+        `self.text_encoder`.
+        All kwargs are forwarded to `self.lora_state_dict`.
+        See [`~loaders.LoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
+        See [`~loaders.LoraLoaderMixin.load_lora_into_transformer`] for more details on how the state dict is loaded
+        into `self.transformer`.
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                See [`~loaders.LoraLoaderMixin.lora_state_dict`].
+            kwargs (`dict`, *optional*):
+                See [`~loaders.LoraLoaderMixin.lora_state_dict`].
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        # if a dict is passed, copy it instead of modifying it inplace
+        if isinstance(pretrained_model_name_or_path_or_dict, dict):
+            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()
+
+        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
+        state_dict, from_diffusers = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
+
+        is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
+        if not is_correct_format:
+            raise ValueError("Invalid LoRA checkpoint.")
+
+        self.load_lora_into_transformer(
+            state_dict,
+            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
+            adapter_name=adapter_name,
+            _pipeline=self,
+            from_diffusers=from_diffusers,
+        )
+
+    @classmethod
+    def lora_state_dict(
+        cls,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, paddle.Tensor]],
+        **kwargs,
+    ):
+        r"""
+        Return state dict for lora weights and the network alphas.
+        <Tip warning={true}>
+        We support loading A1111 formatted LoRA checkpoints in a limited capacity.
+        This function is experimental and might change in the future.
+        </Tip>
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                Can be either:
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+        """
+        # Load the main state dict first which has the LoRA layers for either of
+        # UNet and text encoder or both.
+        from_hf_hub = kwargs.pop("from_hf_hub", FROM_HF_HUB)
+        from_aistudio = kwargs.pop("from_aistudio", FROM_AISTUDIO)
+        cache_dir = kwargs.pop("cache_dir", None)
+        if cache_dir is None:
+            if from_aistudio:
+                cache_dir = None  # TODO, check aistudio cache
+            elif from_hf_hub:
+                cache_dir = DIFFUSERS_CACHE
+            else:
+                cache_dir = PPDIFFUSERS_CACHE
+        from_diffusers = kwargs.pop("from_diffusers", FROM_DIFFUSERS)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        use_auth_token = kwargs.pop("use_auth_token", None)
+        revision = kwargs.pop("revision", None)
+        subfolder = kwargs.pop("subfolder", None)
+        weight_name = kwargs.pop("weight_name", None)
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        allow_pickle = False
+        if use_safetensors is None:
+            use_safetensors = True
+            allow_pickle = True
+
+        user_agent = {
+            "file_type": "attn_procs_weights",
+            "framework": "pytorch" if from_diffusers else "paddle",
+        }
+
+        model_file = None
+        if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+            # Let's first try to load .safetensors weights
+            if (use_safetensors and weight_name is None) or (
+                weight_name is not None and weight_name.endswith(".safetensors")
+            ):
+                try:
+                    model_file = _get_model_file(
+                        pretrained_model_name_or_path_or_dict,
+                        weights_name=(weight_name or TORCH_LORA_WEIGHT_NAME_SAFE)
+                        if from_diffusers
+                        else ((weight_name or PADDLE_LORA_WEIGHT_NAME_SAFE)),
+                        cache_dir=cache_dir,
+                        force_download=force_download,
+                        resume_download=resume_download,
+                        proxies=proxies,
+                        local_files_only=local_files_only,
+                        use_auth_token=use_auth_token,
+                        revision=revision,
+                        subfolder=subfolder,
+                        user_agent=user_agent,
+                        from_aistudio=from_aistudio,
+                        from_hf_hub=from_hf_hub,
+                    )
+                except Exception:
+                    model_file = None
+
+            if model_file is None:
+                model_file = _get_model_file(
+                    pretrained_model_name_or_path_or_dict,
+                    weights_name=(weight_name or TORCH_LORA_WEIGHT_NAME)
+                    if from_diffusers
+                    else ((weight_name or PADDLE_LORA_WEIGHT_NAME)),
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    resume_download=resume_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    use_auth_token=use_auth_token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    user_agent=user_agent,
+                    from_aistudio=from_aistudio,
+                    from_hf_hub=from_hf_hub,
+                )
+
+            assert model_file is not None, "Could not find the model file!"
+            # TODO, check this
+            from ppdiffusers.utils import smart_load
+
+            state_dict = smart_load(model_file, return_is_torch_weight=True)
+            is_torch_weight = state_dict.pop("is_torch_weight", False)
+            if not from_diffusers and is_torch_weight:
+                logger.warning(
+                    "Detect the weight is in diffusers format, but currently, `from_diffusers` is set to `False`. To proceed, we will change the value of `from_diffusers` to `True`!"
+                )
+                from_diffusers = True
+        else:
+            state_dict = pretrained_model_name_or_path_or_dict
+
+        return state_dict, from_diffusers
+
+    @classmethod
+    def load_lora_into_transformer(cls, state_dict, transformer, adapter_name=None, _pipeline=None, from_diffusers=None,):
+        """
+        This will load the LoRA layers specified in `state_dict` into `transformer`.
+        Parameters:
+            state_dict (`dict`):
+                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
+                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
+                encoder lora layers.
+            transformer (`SD3Transformer2DModel`):
+                The Transformer model to load the LoRA layers into.
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+        """
+        if from_diffusers is None:
+            from_diffusers = FROM_DIFFUSERS
+
+        from ppdiffusers.peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict
+
+        keys = list(state_dict.keys())
+
+        transformer_keys = [k for k in keys if k.startswith(cls.transformer_name)]
+        state_dict = {
+            k.replace(f"{cls.transformer_name}.", ""): v for k, v in state_dict.items() if k in transformer_keys
+        }
+
+        if len(state_dict.keys()) > 0:
+            if adapter_name in getattr(transformer, "peft_config", {}):
+                raise ValueError(
+                    f"Adapter name {adapter_name} already in use in the transformer - please select a new adapter name."
+                )
+
+            rank = {}
+            for key, val in state_dict.items():
+                if "lora_B" in key:
+                    rank[key] = val.shape[1]
+
+            lora_config_kwargs = get_peft_kwargs(rank, network_alpha_dict=None, peft_state_dict=state_dict)
+            if "use_dora" in lora_config_kwargs:
+                raise ValueError(
+                    "ppdiffusers.peft does not support dora yet"
+                )
+            lora_config = LoraConfig(**lora_config_kwargs)
+
+            # adapter_name
+            if adapter_name is None:
+                adapter_name = get_adapter_name(transformer)
+
+            inject_adapter_in_model(lora_config, transformer, adapter_name=adapter_name)
+            incompatible_keys = set_peft_model_state_dict(transformer, state_dict, adapter_name)
+
+            if incompatible_keys is not None:
+                # check only for unexpected keys
+                unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
+                if unexpected_keys:
+                    logger.warning(
+                        f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
+                        f" {unexpected_keys}. "
+                    )
+
+            # Unsafe code />
+
+    @classmethod
+    def save_lora_weights(
+        cls,
+        save_directory: Union[str, os.PathLike],
+        transformer_lora_layers: Dict[str, paddle.nn.Layer] = None,
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+        to_diffusers=None,
+    ):
+        r"""
+        Save the LoRA parameters corresponding to the UNet and text encoder.
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to save LoRA parameters to. Will be created if it doesn't exist.
+            transformer_lora_layers (`Dict[str, paddle.nn.Layer]` or `Dict[str, paddletorch.Tensor]`):
+                State dict of the LoRA layers corresponding to the `transformer`.
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful during distributed training and you
+                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+                process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful during distributed training when you need to
+                replace `torch.save` or `paddle.save` with another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional way.
+        """
+        if to_diffusers is None:
+            to_diffusers = TO_DIFFUSERS
+
+        state_dict = {}
+
+        def pack_weights(layers, prefix):
+            layers_weights = layers.state_dict() if isinstance(layers, nn.Layer) else layers
+            layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
+            return layers_state_dict
+
+        if not transformer_lora_layers:
+            raise ValueError("You must pass `transformer_lora_layers`.")
+
+        if transformer_lora_layers:
+            state_dict.update(pack_weights(transformer_lora_layers, cls.transformer_name))
+
+        # Save the model
+        cls.write_lora_layers(
+            state_dict=state_dict,
+            save_directory=save_directory,
+            is_main_process=is_main_process,
+            weight_name=weight_name,
+            save_function=save_function,
+            safe_serialization=safe_serialization,
+            to_diffusers=to_diffusers,  # only change save weights name and save function
+        )
+
+    @staticmethod
+    def write_lora_layers(
+        state_dict: Dict[str, paddle.Tensor],
+        save_directory: str,
+        is_main_process: bool,
+        weight_name: str,
+        save_function: Callable,
+        safe_serialization: bool,
+        to_diffusers=None,
+    ):
+        if to_diffusers is None:
+            to_diffusers = TO_DIFFUSERS
+
+        if os.path.isfile(save_directory):
+            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
+            return
+
+        if save_function is None:
+            if to_diffusers:
+                if not is_torch_available() and not safe_serialization:
+                    safe_serialization = True
+                    logger.warning(
+                        "PyTorch is not installed, and `safe_serialization` is currently set to `False`. "
+                        "To ensure proper model saving, we will automatically set `safe_serialization=True`. "
+                        "If you want to keep `safe_serialization=False`, please make sure PyTorch is installed."
+                    )
+                if safe_serialization:
+                    if is_torch_available():
+                        save_function = partial(torch_safe_save_file, metadata={"format": "pt"})
+                    else:
+                        save_function = partial(np_safe_save_file, metadata={"format": "pt"})
+                else:
+                    save_function = torch.save
+            else:
+                if safe_serialization:
+                    state_dict = {k: np.ascontiguousarray(v) for k, v in state_dict.items()}
+                    save_function = partial(np_safe_save_file, metadata={"format": "pd"})
+                else:
+                    save_function = paddle.save
+
+        os.makedirs(save_directory, exist_ok=True)
+
+        if weight_name is None:
+            if to_diffusers:
+                if safe_serialization:
+                    weight_name = TORCH_LORA_WEIGHT_NAME_SAFE
+                else:
+                    weight_name = TORCH_LORA_WEIGHT_NAME
+            else:
+                if safe_serialization:
+                    weight_name = PADDLE_LORA_WEIGHT_NAME_SAFE
+                else:
+                    weight_name = PADDLE_LORA_WEIGHT_NAME
+        else:
+            if "paddle" in weight_name.lower() or "pdparams" in weight_name.lower():
+                to_diffusers = False
+            elif "torch" in weight_name.lower() or "bin" in weight_name.lower():
+                to_diffusers = True
+
+        save_path = Path(save_directory, weight_name).as_posix()
+        save_function(state_dict, save_path)
+        logger.info(f"Model weights saved in {save_path}")
+
+    def unload_lora_weights(self):
+        """
+        Unloads the LoRA parameters.
+        Examples:
+        ```python
+        >>> # Assuming `pipeline` is already loaded with the LoRA parameters.
+        >>> pipeline.unload_lora_weights()
+        >>> ...
+        ```
+        """
+        transformer = getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer
+        recurse_remove_peft_layers(transformer)
+        if hasattr(transformer, "peft_config"):
+            del transformer.peft_config
+
+    @classmethod
+    def _optionally_disable_offloading(cls, _pipeline):
+        """
+        Optionally removes offloading in case the pipeline has been already sequentially offloaded to CPU.
+        Args:
+            _pipeline (`DiffusionPipeline`):
+                The pipeline to disable offloading for.
+        Returns:
+            tuple:
+                A tuple indicating if `is_model_cpu_offload` or `is_sequential_cpu_offload` is True.
+        """
+        pass

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/loaders/lora_conversion_utils.py

@@ -0,0 +1,283 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import re
+
+from ..utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+def _maybe_map_sgm_blocks_to_diffusers(state_dict, unet_config, delimiter="_", block_slice_pos=5):
+    # 1. get all state_dict_keys
+    all_keys = list(state_dict.keys())
+    sgm_patterns = ["input_blocks", "middle_block", "output_blocks"]
+
+    # 2. check if needs remapping, if not return original dict
+    is_in_sgm_format = False
+    for key in all_keys:
+        if any(p in key for p in sgm_patterns):
+            is_in_sgm_format = True
+            break
+
+    if not is_in_sgm_format:
+        return state_dict
+
+    # 3. Else remap from SGM patterns
+    new_state_dict = {}
+    inner_block_map = ["resnets", "attentions", "upsamplers"]
+
+    # Retrieves # of down, mid and up blocks
+    input_block_ids, middle_block_ids, output_block_ids = set(), set(), set()
+
+    for layer in all_keys:
+        if "text" in layer:
+            new_state_dict[layer] = state_dict.pop(layer)
+        else:
+            layer_id = int(layer.split(delimiter)[:block_slice_pos][-1])
+            if sgm_patterns[0] in layer:
+                input_block_ids.add(layer_id)
+            elif sgm_patterns[1] in layer:
+                middle_block_ids.add(layer_id)
+            elif sgm_patterns[2] in layer:
+                output_block_ids.add(layer_id)
+            else:
+                raise ValueError(f"Checkpoint not supported because layer {layer} not supported.")
+
+    input_blocks = {
+        layer_id: [key for key in state_dict if f"input_blocks{delimiter}{layer_id}" in key]
+        for layer_id in input_block_ids
+    }
+    middle_blocks = {
+        layer_id: [key for key in state_dict if f"middle_block{delimiter}{layer_id}" in key]
+        for layer_id in middle_block_ids
+    }
+    output_blocks = {
+        layer_id: [key for key in state_dict if f"output_blocks{delimiter}{layer_id}" in key]
+        for layer_id in output_block_ids
+    }
+
+    # Rename keys accordingly
+    for i in input_block_ids:
+        block_id = (i - 1) // (unet_config.layers_per_block + 1)
+        layer_in_block_id = (i - 1) % (unet_config.layers_per_block + 1)
+
+        for key in input_blocks[i]:
+            inner_block_id = int(key.split(delimiter)[block_slice_pos])
+            inner_block_key = inner_block_map[inner_block_id] if "op" not in key else "downsamplers"
+            inner_layers_in_block = str(layer_in_block_id) if "op" not in key else "0"
+            new_key = delimiter.join(
+                key.split(delimiter)[: block_slice_pos - 1]
+                + [str(block_id), inner_block_key, inner_layers_in_block]
+                + key.split(delimiter)[block_slice_pos + 1 :]
+            )
+            new_state_dict[new_key] = state_dict.pop(key)
+
+    for i in middle_block_ids:
+        key_part = None
+        if i == 0:
+            key_part = [inner_block_map[0], "0"]
+        elif i == 1:
+            key_part = [inner_block_map[1], "0"]
+        elif i == 2:
+            key_part = [inner_block_map[0], "1"]
+        else:
+            raise ValueError(f"Invalid middle block id {i}.")
+
+        for key in middle_blocks[i]:
+            new_key = delimiter.join(
+                key.split(delimiter)[: block_slice_pos - 1] + key_part + key.split(delimiter)[block_slice_pos:]
+            )
+            new_state_dict[new_key] = state_dict.pop(key)
+
+    for i in output_block_ids:
+        block_id = i // (unet_config.layers_per_block + 1)
+        layer_in_block_id = i % (unet_config.layers_per_block + 1)
+
+        for key in output_blocks[i]:
+            inner_block_id = int(key.split(delimiter)[block_slice_pos])
+            inner_block_key = inner_block_map[inner_block_id]
+            inner_layers_in_block = str(layer_in_block_id) if inner_block_id < 2 else "0"
+            new_key = delimiter.join(
+                key.split(delimiter)[: block_slice_pos - 1]
+                + [str(block_id), inner_block_key, inner_layers_in_block]
+                + key.split(delimiter)[block_slice_pos + 1 :]
+            )
+            new_state_dict[new_key] = state_dict.pop(key)
+
+    if len(state_dict) > 0:
+        raise ValueError("At this point all state dict entries have to be converted.")
+
+    return new_state_dict
+
+
+def _convert_kohya_lora_to_diffusers(state_dict, unet_name="unet", text_encoder_name="text_encoder"):
+    unet_state_dict = {}
+    te_state_dict = {}
+    te2_state_dict = {}
+    network_alphas = {}
+
+    # every down weight has a corresponding up weight and potentially an alpha weight
+    lora_keys = [k for k in state_dict.keys() if k.endswith("lora_down.weight")]
+    for key in lora_keys:
+        lora_name = key.split(".")[0]
+        lora_name_up = lora_name + ".lora_up.weight"
+        lora_name_alpha = lora_name + ".alpha"
+
+        if lora_name.startswith("lora_unet_"):
+            diffusers_name = key.replace("lora_unet_", "").replace("_", ".")
+
+            if "input.blocks" in diffusers_name:
+                diffusers_name = diffusers_name.replace("input.blocks", "down_blocks")
+            else:
+                diffusers_name = diffusers_name.replace("down.blocks", "down_blocks")
+
+            if "middle.block" in diffusers_name:
+                diffusers_name = diffusers_name.replace("middle.block", "mid_block")
+            else:
+                diffusers_name = diffusers_name.replace("mid.block", "mid_block")
+            if "output.blocks" in diffusers_name:
+                diffusers_name = diffusers_name.replace("output.blocks", "up_blocks")
+            else:
+                diffusers_name = diffusers_name.replace("up.blocks", "up_blocks")
+
+            diffusers_name = diffusers_name.replace("transformer.blocks", "transformer_blocks")
+            diffusers_name = diffusers_name.replace("to.q.lora", "to_q_lora")
+            diffusers_name = diffusers_name.replace("to.k.lora", "to_k_lora")
+            diffusers_name = diffusers_name.replace("to.v.lora", "to_v_lora")
+            diffusers_name = diffusers_name.replace("to.out.0.lora", "to_out_lora")
+            diffusers_name = diffusers_name.replace("proj.in", "proj_in")
+            diffusers_name = diffusers_name.replace("proj.out", "proj_out")
+            diffusers_name = diffusers_name.replace("emb.layers", "time_emb_proj")
+
+            # SDXL specificity.
+            if "emb" in diffusers_name and "time.emb.proj" not in diffusers_name:
+                pattern = r"\.\d+(?=\D*$)"
+                diffusers_name = re.sub(pattern, "", diffusers_name, count=1)
+            if ".in." in diffusers_name:
+                diffusers_name = diffusers_name.replace("in.layers.2", "conv1")
+            if ".out." in diffusers_name:
+                diffusers_name = diffusers_name.replace("out.layers.3", "conv2")
+            if "downsamplers" in diffusers_name or "upsamplers" in diffusers_name:
+                diffusers_name = diffusers_name.replace("op", "conv")
+            if "skip" in diffusers_name:
+                diffusers_name = diffusers_name.replace("skip.connection", "conv_shortcut")
+
+            # LyCORIS specificity.
+            if "time.emb.proj" in diffusers_name:
+                diffusers_name = diffusers_name.replace("time.emb.proj", "time_emb_proj")
+            if "conv.shortcut" in diffusers_name:
+                diffusers_name = diffusers_name.replace("conv.shortcut", "conv_shortcut")
+
+            # General coverage.
+            if "transformer_blocks" in diffusers_name:
+                if "attn1" in diffusers_name or "attn2" in diffusers_name:
+                    diffusers_name = diffusers_name.replace("attn1", "attn1.processor")
+                    diffusers_name = diffusers_name.replace("attn2", "attn2.processor")
+                    unet_state_dict[diffusers_name] = state_dict.pop(key)
+                    unet_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+                elif "ff" in diffusers_name:
+                    unet_state_dict[diffusers_name] = state_dict.pop(key)
+                    unet_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+            elif any(key in diffusers_name for key in ("proj_in", "proj_out")):
+                unet_state_dict[diffusers_name] = state_dict.pop(key)
+                unet_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+            else:
+                unet_state_dict[diffusers_name] = state_dict.pop(key)
+                unet_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+
+        elif lora_name.startswith("lora_te_"):
+            diffusers_name = key.replace("lora_te_", "").replace("_", ".")
+            diffusers_name = diffusers_name.replace("text.model", "text_model")
+            diffusers_name = diffusers_name.replace("self.attn", "self_attn")
+            diffusers_name = diffusers_name.replace("q.proj.lora", "to_q_lora")
+            diffusers_name = diffusers_name.replace("k.proj.lora", "to_k_lora")
+            diffusers_name = diffusers_name.replace("v.proj.lora", "to_v_lora")
+            diffusers_name = diffusers_name.replace("out.proj.lora", "to_out_lora")
+            if "self_attn" in diffusers_name:
+                te_state_dict[diffusers_name] = state_dict.pop(key)
+                te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+            elif "mlp" in diffusers_name:
+                # Be aware that this is the new diffusers convention and the rest of the code might
+                # not utilize it yet.
+                diffusers_name = diffusers_name.replace(".lora.", ".lora_linear_layer.")
+                te_state_dict[diffusers_name] = state_dict.pop(key)
+                te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+
+        # (sayakpaul): Duplicate code. Needs to be cleaned.
+        elif lora_name.startswith("lora_te1_"):
+            diffusers_name = key.replace("lora_te1_", "").replace("_", ".")
+            diffusers_name = diffusers_name.replace("text.model", "text_model")
+            diffusers_name = diffusers_name.replace("self.attn", "self_attn")
+            diffusers_name = diffusers_name.replace("q.proj.lora", "to_q_lora")
+            diffusers_name = diffusers_name.replace("k.proj.lora", "to_k_lora")
+            diffusers_name = diffusers_name.replace("v.proj.lora", "to_v_lora")
+            diffusers_name = diffusers_name.replace("out.proj.lora", "to_out_lora")
+            if "self_attn" in diffusers_name:
+                te_state_dict[diffusers_name] = state_dict.pop(key)
+                te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+            elif "mlp" in diffusers_name:
+                # Be aware that this is the new diffusers convention and the rest of the code might
+                # not utilize it yet.
+                diffusers_name = diffusers_name.replace(".lora.", ".lora_linear_layer.")
+                te_state_dict[diffusers_name] = state_dict.pop(key)
+                te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+
+        # (sayakpaul): Duplicate code. Needs to be cleaned.
+        elif lora_name.startswith("lora_te2_"):
+            diffusers_name = key.replace("lora_te2_", "").replace("_", ".")
+            diffusers_name = diffusers_name.replace("text.model", "text_model")
+            diffusers_name = diffusers_name.replace("self.attn", "self_attn")
+            diffusers_name = diffusers_name.replace("q.proj.lora", "to_q_lora")
+            diffusers_name = diffusers_name.replace("k.proj.lora", "to_k_lora")
+            diffusers_name = diffusers_name.replace("v.proj.lora", "to_v_lora")
+            diffusers_name = diffusers_name.replace("out.proj.lora", "to_out_lora")
+            if "self_attn" in diffusers_name:
+                te2_state_dict[diffusers_name] = state_dict.pop(key)
+                te2_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+            elif "mlp" in diffusers_name:
+                # Be aware that this is the new diffusers convention and the rest of the code might
+                # not utilize it yet.
+                diffusers_name = diffusers_name.replace(".lora.", ".lora_linear_layer.")
+                te2_state_dict[diffusers_name] = state_dict.pop(key)
+                te2_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+
+        # Rename the alphas so that they can be mapped appropriately.
+        if lora_name_alpha in state_dict:
+            alpha = state_dict.pop(lora_name_alpha).item()
+            if lora_name_alpha.startswith("lora_unet_"):
+                prefix = "unet."
+            elif lora_name_alpha.startswith(("lora_te_", "lora_te1_")):
+                prefix = "text_encoder."
+            else:
+                prefix = "text_encoder_2."
+            new_name = prefix + diffusers_name.split(".lora.")[0] + ".alpha"
+            network_alphas.update({new_name: alpha})
+
+    if len(state_dict) > 0:
+        raise ValueError(f"The following keys have not been correctly be renamed: \n\n {', '.join(state_dict.keys())}")
+
+    logger.info("Kohya-style checkpoint detected.")
+    unet_state_dict = {f"{unet_name}.{module_name}": params for module_name, params in unet_state_dict.items()}
+    te_state_dict = {f"{text_encoder_name}.{module_name}": params for module_name, params in te_state_dict.items()}
+    te2_state_dict = (
+        {f"text_encoder_2.{module_name}": params for module_name, params in te2_state_dict.items()}
+        if len(te2_state_dict) > 0
+        else None
+    )
+    if te2_state_dict is not None:
+        te_state_dict.update(te2_state_dict)
+
+    new_state_dict = {**unet_state_dict, **te_state_dict}
+    return new_state_dict, network_alphas

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/loaders/single_file.py

@@ -0,0 +1,766 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import copy
+import os
+from io import BytesIO
+from pathlib import Path
+
+import paddle
+import requests
+from huggingface_hub import hf_hub_download
+from huggingface_hub.file_download import _request_wrapper, hf_raise_for_status
+
+from ..utils import (
+    BACKENDS_MAPPING,
+    DIFFUSERS_CACHE,
+    HF_HUB_OFFLINE,
+    PPDIFFUSERS_CACHE,
+    deprecate,
+    is_omegaconf_available,
+    logging,
+    ppdiffusers_url_download,
+    smart_load,
+)
+
+try:
+    from paddlenlp.transformers.model_utils import no_init_weights
+except ImportError:
+    from ..utils.paddle_utils import no_init_weights
+
+from ..models.modeling_utils import ContextManagers, faster_set_state_dict
+
+
+def http_file_name(
+    url: str,
+    *,
+    proxies=None,
+    headers=None,
+    timeout=10.0,
+    max_retries=0,
+):
+    """
+    Get a remote file name.
+    """
+    headers = copy.deepcopy(headers) or {}
+    r = _request_wrapper(
+        method="GET",
+        url=url,
+        stream=True,
+        proxies=proxies,
+        headers=headers,
+        timeout=timeout,
+        # max_retries=max_retries,
+    )
+    hf_raise_for_status(r)
+    displayed_name = url.split("/")[-1]
+    content_disposition = r.headers.get("Content-Disposition")
+    if content_disposition is not None and "filename=" in content_disposition:
+        # Means file is on CDN
+        displayed_name = content_disposition.split("filename=")[-1]
+    return displayed_name
+
+
+logger = logging.get_logger(__name__)
+
+
+class FromSingleFileMixin:
+    """
+    Load model weights saved in the `.ckpt` format into a [`DiffusionPipeline`].
+    """
+
+    @classmethod
+    def from_ckpt(cls, *args, **kwargs):
+        deprecation_message = "The function `from_ckpt` is deprecated in favor of `from_single_file` and will be removed in ppdiffusers v.0.21. Please make sure to use `StableDiffusionPipeline.from_single_file(...)` instead."
+        deprecate("from_ckpt", "0.21.0", deprecation_message, standard_warn=False)
+        return cls.from_single_file(*args, **kwargs)
+
+    @classmethod
+    def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
+        r"""
+        Instantiate a [`DiffusionPipeline`] from pretrained pipeline weights saved in the `.ckpt` or `.safetensors`
+        format. The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+        Parameters:
+            pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
+                Can be either:
+                    - A link to the `.ckpt` file (for example
+                      `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
+                    - A path to a *file* containing all pipeline weights.
+            paddle_dtype (`str` or `paddle.dtype`, *optional*):
+                Override the default `paddle.dtype` and load the model with another dtype. If `"auto"` is passed, the
+                dtype is automatically derived from the model's weights.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            use_auth_token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            use_safetensors (`bool`, *optional*, defaults to `None`):
+                If set to `None`, the safetensors weights are downloaded if they're available **and** if the
+                safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
+                weights. If set to `False`, safetensors weights are not loaded.
+            extract_ema (`bool`, *optional*, defaults to `False`):
+                Whether to extract the EMA weights or not. Pass `True` to extract the EMA weights which usually yield
+                higher quality images for inference. Non-EMA weights are usually better for continuing finetuning.
+            upcast_attention (`bool`, *optional*, defaults to `None`):
+                Whether the attention computation should always be upcasted.
+            image_size (`int`, *optional*, defaults to 512):
+                The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
+                Diffusion v2 base model. Use 768 for Stable Diffusion v2.
+            prediction_type (`str`, *optional*):
+                The prediction type the model was trained on. Use `'epsilon'` for all Stable Diffusion v1 models and
+                the Stable Diffusion v2 base model. Use `'v_prediction'` for Stable Diffusion v2.
+            num_in_channels (`int`, *optional*, defaults to `None`):
+                The number of input channels. If `None`, it is automatically inferred.
+            scheduler_type (`str`, *optional*, defaults to `"pndm"`):
+                Type of scheduler to use. Should be one of `["pndm", "lms", "heun", "euler", "euler-ancestral", "dpm",
+                "ddim"]`.
+            load_safety_checker (`bool`, *optional*, defaults to `True`):
+                Whether to load the safety checker or not.
+            text_encoder ([`~transformers.CLIPTextModel`], *optional*, defaults to `None`):
+                An instance of `CLIPTextModel` to use, specifically the
+                [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. If this
+                parameter is `None`, the function loads a new instance of `CLIPTextModel` by itself if needed.
+            vae (`AutoencoderKL`, *optional*, defaults to `None`):
+                Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. If
+                this parameter is `None`, the function will load a new instance of [CLIP] by itself, if needed.
+            tokenizer ([`~transformers.CLIPTokenizer`], *optional*, defaults to `None`):
+                An instance of `CLIPTokenizer` to use. If this parameter is `None`, the function loads a new instance
+                of `CLIPTokenizer` by itself if needed.
+            original_config_file (`str`):
+                Path to `.yaml` config file corresponding to the original architecture. If `None`, will be
+                automatically inferred by looking for a key that only exists in SD2.0 models.
+            kwargs (remaining dictionary of keyword arguments, *optional*):
+                Can be used to overwrite load and saveable variables (for example the pipeline components of the
+                specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
+                method. See example below for more information.
+
+        Examples:
+
+        ```py
+        >>> from ppdiffusers import StableDiffusionPipeline
+
+        >>> # Download pipeline from huggingface.co and cache.
+        >>> pipeline = StableDiffusionPipeline.from_single_file(
+        ...     "https://huggingface.co/WarriorMama777/OrangeMixs/blob/main/Models/AbyssOrangeMix/AbyssOrangeMix.safetensors"
+        ... )
+
+        >>> # Download pipeline from local file
+        >>> # file is downloaded under ./v1-5-pruned-emaonly.ckpt
+        >>> pipeline = StableDiffusionPipeline.from_single_file("./v1-5-pruned-emaonly")
+
+        >>> # Enable float16
+        >>> pipeline = StableDiffusionPipeline.from_single_file(
+        ...     "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.ckpt",
+        ...     paddle_dtype=paddle.float16,
+        ... )
+        ```
+        """
+        # import here to avoid circular dependency
+        from ..pipelines.stable_diffusion.convert_from_ckpt import (
+            download_from_original_stable_diffusion_ckpt,
+        )
+
+        from_hf_hub = any(p in pretrained_model_link_or_path for p in ["huggingface.co", "hf.co", "hf-mirror"])
+        cache_dir = (
+            kwargs.pop("cache_dir", DIFFUSERS_CACHE) if from_hf_hub else kwargs.pop("cache_dir", PPDIFFUSERS_CACHE)
+        )
+        original_config_file = kwargs.pop("original_config_file", None)
+        config_files = kwargs.pop("config_files", None)
+        resume_download = kwargs.pop("resume_download", False)
+        force_download = kwargs.pop("force_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
+        revision = kwargs.pop("revision", None)
+        extract_ema = kwargs.pop("extract_ema", False)
+        image_size = kwargs.pop("image_size", None)
+        scheduler_type = kwargs.pop("scheduler_type", "pndm")
+        num_in_channels = kwargs.pop("num_in_channels", None)
+        upcast_attention = kwargs.pop("upcast_attention", None)
+        load_safety_checker = kwargs.pop("load_safety_checker", False)
+        prediction_type = kwargs.pop("prediction_type", None)
+        text_encoder = kwargs.pop("text_encoder", None)
+        vae = kwargs.pop("vae", None)
+        controlnet = kwargs.pop("controlnet", None)
+        adapter = kwargs.pop("adapter", None)
+        tokenizer = kwargs.pop("tokenizer", None)
+
+        paddle_dtype = kwargs.pop("paddle_dtype", None)
+
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        pipeline_name = cls.__name__
+        file_extension = pretrained_model_link_or_path.rsplit(".", 1)[-1]
+        from_safetensors = file_extension == "safetensors"
+
+        if from_safetensors and use_safetensors is False:
+            raise ValueError("Make sure to install `safetensors` with `pip install safetensors`.")
+
+        # TODO: For now we only support stable diffusion
+        stable_unclip = None
+        model_type = None
+
+        if pipeline_name in [
+            "StableDiffusionControlNetPipeline",
+            "StableDiffusionControlNetImg2ImgPipeline",
+            "StableDiffusionControlNetInpaintPipeline",
+        ]:
+            from ..models.controlnet import ControlNetModel
+            from ..pipelines.controlnet.multicontrolnet import MultiControlNetModel
+
+            #  list/tuple or a single instance of ControlNetModel or MultiControlNetModel
+            if not (
+                isinstance(controlnet, (ControlNetModel, MultiControlNetModel))
+                or isinstance(controlnet, (list, tuple))
+                and isinstance(controlnet[0], ControlNetModel)
+            ):
+                raise ValueError("ControlNet needs to be passed if loading from ControlNet pipeline.")
+        elif "StableDiffusion" in pipeline_name:
+            # Model type will be inferred from the checkpoint.
+            pass
+        elif pipeline_name == "StableUnCLIPPipeline":
+            model_type = "FrozenOpenCLIPEmbedder"
+            stable_unclip = "txt2img"
+        elif pipeline_name == "StableUnCLIPImg2ImgPipeline":
+            model_type = "FrozenOpenCLIPEmbedder"
+            stable_unclip = "img2img"
+        elif pipeline_name == "PaintByExamplePipeline":
+            model_type = "PaintByExample"
+        elif pipeline_name == "LDMTextToImagePipeline":
+            model_type = "LDMTextToImage"
+        else:
+            raise ValueError(f"Unhandled pipeline class: {pipeline_name}")
+
+        pretrained_model_link_or_path = str(pretrained_model_link_or_path)
+        if os.path.isfile(pretrained_model_link_or_path):
+            checkpoint_path = pretrained_model_link_or_path
+        elif pretrained_model_link_or_path.startswith("http://") or pretrained_model_link_or_path.startswith(
+            "https://"
+        ):
+            if from_hf_hub:
+                # remove huggingface url
+                has_valid_url_prefix = False
+                valid_url_prefixes = [
+                    "https://huggingface.co/",
+                    "huggingface.co/",
+                    "hf.co/",
+                    "https://hf.co/",
+                    "hf-mirror.com/",
+                    "https://hf-mirror.com/",
+                ]
+                for prefix in valid_url_prefixes:
+                    if pretrained_model_link_or_path.startswith(prefix):
+                        pretrained_model_link_or_path = pretrained_model_link_or_path[len(prefix) :]
+                        has_valid_url_prefix = True
+
+                # Code based on diffusers.pipelines.pipeline_utils.DiffusionPipeline.from_pretrained
+                ckpt_path = Path(pretrained_model_link_or_path)
+                if not ckpt_path.is_file():
+                    if not has_valid_url_prefix:
+                        raise ValueError(
+                            f"The provided path is either not a file or a valid huggingface URL was not provided. Valid URLs begin with {', '.join(valid_url_prefixes)}"
+                        )
+
+                    # get repo_id and (potentially nested) file path of ckpt in repo
+                    repo_id = "/".join(ckpt_path.parts[:2])
+                    file_path = "/".join(ckpt_path.parts[2:])
+
+                    if file_path.startswith("blob/"):
+                        file_path = file_path[len("blob/") :]
+
+                    if file_path.startswith("main/"):
+                        file_path = file_path[len("main/") :]
+
+                    checkpoint_path = hf_hub_download(
+                        repo_id,
+                        filename=file_path,
+                        cache_dir=cache_dir,
+                        resume_download=resume_download,
+                        proxies=proxies,
+                        local_files_only=local_files_only,
+                        token=use_auth_token,
+                        revision=revision,
+                        force_download=force_download,
+                    )
+                else:
+                    checkpoint_path = ckpt_path
+            else:
+                checkpoint_path = ppdiffusers_url_download(
+                    pretrained_model_link_or_path,
+                    cache_dir=cache_dir,
+                    filename=http_file_name(pretrained_model_link_or_path).strip('"'),
+                    force_download=force_download,
+                    resume_download=resume_download,
+                )
+        else:
+            checkpoint_path = pretrained_model_link_or_path
+
+        pipe = download_from_original_stable_diffusion_ckpt(
+            checkpoint_path,
+            pipeline_class=cls,
+            model_type=model_type,
+            stable_unclip=stable_unclip,
+            controlnet=controlnet,
+            adapter=adapter,
+            from_safetensors=from_safetensors,
+            extract_ema=extract_ema,
+            image_size=image_size,
+            scheduler_type=scheduler_type,
+            num_in_channels=num_in_channels,
+            upcast_attention=upcast_attention,
+            load_safety_checker=load_safety_checker,
+            prediction_type=prediction_type,
+            paddle_dtype=paddle_dtype,
+            text_encoder=text_encoder,
+            vae=vae,
+            tokenizer=tokenizer,
+            original_config_file=original_config_file,
+            config_files=config_files,
+            local_files_only=local_files_only,
+        )
+
+        return pipe
+
+
+class FromOriginalVAEMixin:
+    """
+    Load pretrained ControlNet weights saved in the `.ckpt` or `.safetensors` format into an [`AutoencoderKL`].
+    """
+
+    @classmethod
+    def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
+        r"""
+        Instantiate a [`AutoencoderKL`] from pretrained ControlNet weights saved in the original `.ckpt` or
+        `.safetensors` format. The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+        Parameters:
+            pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
+                Can be either:
+                    - A link to the `.ckpt` file (for example
+                      `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
+                    - A path to a *file* containing all pipeline weights.
+            paddle_dtype (`str` or `paddle.dtype`, *optional*):
+                Override the default `paddle.dtype` and load the model with another dtype. If `"auto"` is passed, the
+                dtype is automatically derived from the model's weights.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to True, the model
+                won't be downloaded from the Hub.
+            use_auth_token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            image_size (`int`, *optional*, defaults to 512):
+                The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
+                Diffusion v2 base model. Use 768 for Stable Diffusion v2.
+            use_safetensors (`bool`, *optional*, defaults to `None`):
+                If set to `None`, the safetensors weights are downloaded if they're available **and** if the
+                safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
+                weights. If set to `False`, safetensors weights are not loaded.
+            upcast_attention (`bool`, *optional*, defaults to `None`):
+                Whether the attention computation should always be upcasted.
+            scaling_factor (`float`, *optional*, defaults to 0.18215):
+                The component-wise standard deviation of the trained latent space computed using the first batch of the
+                training set. This is used to scale the latent space to have unit variance when training the diffusion
+                model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+                diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z
+                = 1 / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution
+                Image Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+            kwargs (remaining dictionary of keyword arguments, *optional*):
+                Can be used to overwrite load and saveable variables (for example the pipeline components of the
+                specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
+                method. See example below for more information.
+
+        <Tip warning={true}>
+
+            Make sure to pass both `image_size` and `scaling_factor` to `from_single_file()` if you're loading
+            a VAE from SDXL or a Stable Diffusion v2 model or higher.
+
+        </Tip>
+
+        Examples:
+
+        ```py
+        from ppdiffusers import AutoencoderKL
+
+        url = "https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.safetensors"  # can also be local file
+        model = AutoencoderKL.from_single_file(url)
+        ```
+        """
+        from_hf_hub = any(p in pretrained_model_link_or_path for p in ["huggingface.co", "hf.co", "hf-mirror"])
+        cache_dir = (
+            kwargs.pop("cache_dir", DIFFUSERS_CACHE) if from_hf_hub else kwargs.pop("cache_dir", PPDIFFUSERS_CACHE)
+        )
+
+        if not is_omegaconf_available():
+            raise ValueError(BACKENDS_MAPPING["omegaconf"][1])
+
+        from omegaconf import OmegaConf
+
+        from ..models import AutoencoderKL
+
+        # import here to avoid circular dependency
+        from ..pipelines.stable_diffusion.convert_from_ckpt import (
+            convert_diffusers_vae_unet_to_ppdiffusers,
+            convert_ldm_vae_checkpoint,
+            create_vae_diffusers_config,
+        )
+
+        config_file = kwargs.pop("config_file", None)
+        resume_download = kwargs.pop("resume_download", False)
+        force_download = kwargs.pop("force_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
+        revision = kwargs.pop("revision", None)
+        image_size = kwargs.pop("image_size", None)
+        scaling_factor = kwargs.pop("scaling_factor", None)
+        kwargs.pop("upcast_attention", None)
+
+        paddle_dtype = kwargs.pop("paddle_dtype", None)
+
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        file_extension = pretrained_model_link_or_path.rsplit(".", 1)[-1]
+        from_safetensors = file_extension == "safetensors"
+
+        if from_safetensors and use_safetensors is False:
+            raise ValueError("Make sure to install `safetensors` with `pip install safetensors`.")
+
+        pretrained_model_link_or_path = str(pretrained_model_link_or_path)
+        if os.path.isfile(pretrained_model_link_or_path):
+            checkpoint_path = pretrained_model_link_or_path
+        elif pretrained_model_link_or_path.startswith("http://") or pretrained_model_link_or_path.startswith(
+            "https://"
+        ):
+            if from_hf_hub:
+                # remove huggingface url
+                has_valid_url_prefix = False
+                valid_url_prefixes = [
+                    "https://huggingface.co/",
+                    "huggingface.co/",
+                    "hf.co/",
+                    "https://hf.co/",
+                    "hf-mirror.com/",
+                    "https://hf-mirror.com/",
+                ]
+                for prefix in valid_url_prefixes:
+                    if pretrained_model_link_or_path.startswith(prefix):
+                        pretrained_model_link_or_path = pretrained_model_link_or_path[len(prefix) :]
+                        has_valid_url_prefix = True
+
+                # Code based on diffusers.pipelines.pipeline_utils.DiffusionPipeline.from_pretrained
+                ckpt_path = Path(pretrained_model_link_or_path)
+                if not ckpt_path.is_file():
+                    if not has_valid_url_prefix:
+                        raise ValueError(
+                            f"The provided path is either not a file or a valid huggingface URL was not provided. Valid URLs begin with {', '.join(valid_url_prefixes)}"
+                        )
+                    # get repo_id and (potentially nested) file path of ckpt in repo
+                    repo_id = "/".join(ckpt_path.parts[:2])
+                    file_path = "/".join(ckpt_path.parts[2:])
+
+                    if file_path.startswith("blob/"):
+                        file_path = file_path[len("blob/") :]
+
+                    if file_path.startswith("main/"):
+                        file_path = file_path[len("main/") :]
+
+                    checkpoint_path = hf_hub_download(
+                        repo_id,
+                        filename=file_path,
+                        cache_dir=cache_dir,
+                        resume_download=resume_download,
+                        proxies=proxies,
+                        local_files_only=local_files_only,
+                        token=use_auth_token,
+                        revision=revision,
+                        force_download=force_download,
+                    )
+                else:
+                    checkpoint_path = ckpt_path
+
+            else:
+                checkpoint_path = ppdiffusers_url_download(
+                    pretrained_model_link_or_path,
+                    cache_dir=cache_dir,
+                    filename=http_file_name(pretrained_model_link_or_path).strip('"'),
+                    force_download=force_download,
+                    resume_download=resume_download,
+                )
+        else:
+            checkpoint_path = pretrained_model_link_or_path
+        checkpoint = smart_load(checkpoint_path, return_numpy=True)
+
+        if "state_dict" in checkpoint:
+            checkpoint = checkpoint["state_dict"]
+
+        if config_file is None:
+            config_url = "https://paddlenlp.bj.bcebos.com/models/community/junnyu/develop/v1-inference.yaml"
+            config_file = BytesIO(requests.get(config_url).content)
+
+        original_config = OmegaConf.load(config_file)
+
+        # default to sd-v1-5
+        image_size = image_size or 512
+
+        vae_config = create_vae_diffusers_config(original_config, image_size=image_size)
+        converted_vae_checkpoint = convert_ldm_vae_checkpoint(checkpoint, vae_config)
+
+        if scaling_factor is None:
+            if (
+                "model" in original_config
+                and "params" in original_config.model
+                and "scale_factor" in original_config.model.params
+            ):
+                vae_scaling_factor = original_config.model.params.scale_factor
+            else:
+                vae_scaling_factor = 0.18215  # default SD scaling factor
+
+        vae_config["scaling_factor"] = vae_scaling_factor
+
+        init_contexts = []
+        init_contexts.append(paddle.dtype_guard(paddle.float32))
+        init_contexts.append(no_init_weights(_enable=True))
+        if hasattr(paddle, "LazyGuard"):
+            init_contexts.append(paddle.LazyGuard())
+        with ContextManagers(init_contexts):
+            vae = AutoencoderKL(**vae_config)
+
+        # we must transpose linear layer
+        faster_set_state_dict(vae, convert_diffusers_vae_unet_to_ppdiffusers(vae, converted_vae_checkpoint))
+
+        if paddle_dtype is not None:
+            vae.to(paddle_dtype=paddle_dtype)
+
+        return vae
+
+
+class FromOriginalControlnetMixin:
+    """
+    Load pretrained ControlNet weights saved in the `.ckpt` or `.safetensors` format into a [`ControlNetModel`].
+    """
+
+    @classmethod
+    def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
+        r"""
+        Instantiate a [`ControlNetModel`] from pretrained ControlNet weights saved in the original `.ckpt` or
+        `.safetensors` format. The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+        Parameters:
+            pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
+                Can be either:
+                    - A link to the `.ckpt` file (for example
+                      `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
+                    - A path to a *file* containing all pipeline weights.
+            paddle_dtype (`str` or `paddle.dtype`, *optional*):
+                Override the default `paddle.dtype` and load the model with another dtype. If `"auto"` is passed, the
+                dtype is automatically derived from the model's weights.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to True, the model
+                won't be downloaded from the Hub.
+            use_auth_token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            use_safetensors (`bool`, *optional*, defaults to `None`):
+                If set to `None`, the safetensors weights are downloaded if they're available **and** if the
+                safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
+                weights. If set to `False`, safetensors weights are not loaded.
+            image_size (`int`, *optional*, defaults to 512):
+                The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
+                Diffusion v2 base model. Use 768 for Stable Diffusion v2.
+            upcast_attention (`bool`, *optional*, defaults to `None`):
+                Whether the attention computation should always be upcasted.
+            kwargs (remaining dictionary of keyword arguments, *optional*):
+                Can be used to overwrite load and saveable variables (for example the pipeline components of the
+                specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
+                method. See example below for more information.
+
+        Examples:
+
+        ```py
+        from ppdiffusers import StableDiffusionControlNetPipeline, ControlNetModel
+
+        url = "https://huggingface.co/lllyasviel/ControlNet-v1-1/blob/main/control_v11p_sd15_canny.pth"  # can also be a local path
+        model = ControlNetModel.from_single_file(url)
+
+        url = "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned.safetensors"  # can also be a local path
+        pipe = StableDiffusionControlNetPipeline.from_single_file(url, controlnet=controlnet)
+        ```
+        """
+        # import here to avoid circular dependency
+        from ..pipelines.stable_diffusion.convert_from_ckpt import (
+            download_controlnet_from_original_ckpt,
+        )
+
+        from_hf_hub = any(p in pretrained_model_link_or_path for p in ["huggingface.co", "hf.co", "hf-mirror"])
+        cache_dir = (
+            kwargs.pop("cache_dir", DIFFUSERS_CACHE) if from_hf_hub else kwargs.pop("cache_dir", PPDIFFUSERS_CACHE)
+        )
+
+        config_file = kwargs.pop("config_file", None)
+        resume_download = kwargs.pop("resume_download", False)
+        force_download = kwargs.pop("force_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
+        num_in_channels = kwargs.pop("num_in_channels", None)
+        use_linear_projection = kwargs.pop("use_linear_projection", None)
+        revision = kwargs.pop("revision", None)
+        extract_ema = kwargs.pop("extract_ema", False)
+        image_size = kwargs.pop("image_size", None)
+        upcast_attention = kwargs.pop("upcast_attention", None)
+
+        paddle_dtype = kwargs.pop("paddle_dtype", None)
+
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        file_extension = pretrained_model_link_or_path.rsplit(".", 1)[-1]
+        from_safetensors = file_extension == "safetensors"
+
+        if from_safetensors and use_safetensors is False:
+            raise ValueError("Make sure to install `safetensors` with `pip install safetensors`.")
+
+        pretrained_model_link_or_path = str(pretrained_model_link_or_path)
+        if os.path.isfile(pretrained_model_link_or_path):
+            checkpoint_path = pretrained_model_link_or_path
+        elif pretrained_model_link_or_path.startswith("http://") or pretrained_model_link_or_path.startswith(
+            "https://"
+        ):
+            if from_hf_hub:
+                # remove huggingface url
+                has_valid_url_prefix = False
+                valid_url_prefixes = [
+                    "https://huggingface.co/",
+                    "huggingface.co/",
+                    "hf.co/",
+                    "https://hf.co/",
+                    "hf-mirror.com/",
+                    "https://hf-mirror.com/",
+                ]
+                for prefix in valid_url_prefixes:
+                    if pretrained_model_link_or_path.startswith(prefix):
+                        pretrained_model_link_or_path = pretrained_model_link_or_path[len(prefix) :]
+                        has_valid_url_prefix = True
+
+                # Code based on diffusers.pipelines.pipeline_utils.DiffusionPipeline.from_pretrained
+                ckpt_path = Path(pretrained_model_link_or_path)
+                if not ckpt_path.is_file():
+                    if not has_valid_url_prefix:
+                        raise ValueError(
+                            f"The provided path is either not a file or a valid huggingface URL was not provided. Valid URLs begin with {', '.join(valid_url_prefixes)}"
+                        )
+                    # get repo_id and (potentially nested) file path of ckpt in repo
+                    repo_id = "/".join(ckpt_path.parts[:2])
+                    file_path = "/".join(ckpt_path.parts[2:])
+
+                    if file_path.startswith("blob/"):
+                        file_path = file_path[len("blob/") :]
+
+                    if file_path.startswith("main/"):
+                        file_path = file_path[len("main/") :]
+
+                    checkpoint_path = hf_hub_download(
+                        repo_id,
+                        filename=file_path,
+                        cache_dir=cache_dir,
+                        resume_download=resume_download,
+                        proxies=proxies,
+                        local_files_only=local_files_only,
+                        token=use_auth_token,
+                        revision=revision,
+                        force_download=force_download,
+                    )
+                else:
+                    checkpoint_path = ckpt_path
+
+            else:
+                checkpoint_path = ppdiffusers_url_download(
+                    pretrained_model_link_or_path,
+                    cache_dir=cache_dir,
+                    filename=http_file_name(pretrained_model_link_or_path).strip('"'),
+                    force_download=force_download,
+                    resume_download=resume_download,
+                )
+        else:
+            checkpoint_path = pretrained_model_link_or_path
+
+        if config_file is None:
+            config_url = "https://paddlenlp.bj.bcebos.com/models/community/junnyu/develop/cldm_v15.yaml"
+            config_file = BytesIO(requests.get(config_url).content)
+
+        image_size = image_size or 512
+
+        controlnet = download_controlnet_from_original_ckpt(
+            checkpoint_path,
+            original_config_file=config_file,
+            image_size=image_size,
+            extract_ema=extract_ema,
+            num_in_channels=num_in_channels,
+            upcast_attention=upcast_attention,
+            from_safetensors=from_safetensors,
+            use_linear_projection=use_linear_projection,
+        )
+
+        if paddle_dtype is not None:
+            controlnet.to(paddle_dtype=paddle_dtype)
+
+        return controlnet
+
+
+FromCkptMixin = FromSingleFileMixin

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/loaders/textual_inversion.py

@@ -0,0 +1,449 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Dict, List, Optional, Union
+
+import paddle
+
+from ..utils import (
+    DIFFUSERS_CACHE,
+    FROM_AISTUDIO,
+    FROM_DIFFUSERS,
+    FROM_HF_HUB,
+    HF_HUB_OFFLINE,
+    PPDIFFUSERS_CACHE,
+    _get_model_file,
+    is_paddlenlp_available,
+    logging,
+)
+
+if is_paddlenlp_available():
+    from paddlenlp.transformers import PretrainedModel, PretrainedTokenizer
+
+logger = logging.get_logger(__name__)
+from ..models.modeling_utils import load_state_dict
+
+TORCH_TEXT_INVERSION_NAME = "learned_embeds.bin"
+PADDLE_TEXT_INVERSION_NAME = "learned_embeds.pdparams"
+TEXT_INVERSION_NAME_SAFE = "learned_embeds.safetensors"
+
+
+def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs):
+    from_hf_hub = kwargs.pop("from_hf_hub", FROM_HF_HUB)
+    from_aistudio = kwargs.pop("from_aistudio", FROM_AISTUDIO)
+    from_diffusers = kwargs.pop("from_diffusers", FROM_DIFFUSERS)
+
+    cache_dir = kwargs.pop("cache_dir", None)
+    if cache_dir is None:
+        if from_aistudio:
+            cache_dir = None  # TODO, check aistudio cache
+        elif from_hf_hub:
+            cache_dir = DIFFUSERS_CACHE
+        else:
+            cache_dir = PPDIFFUSERS_CACHE
+    force_download = kwargs.pop("force_download", False)
+    resume_download = kwargs.pop("resume_download", False)
+    proxies = kwargs.pop("proxies", None)
+    local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+    use_auth_token = kwargs.pop("use_auth_token", None)
+    revision = kwargs.pop("revision", None)
+    subfolder = kwargs.pop("subfolder", None)
+    weight_name = kwargs.pop("weight_name", None)
+    use_safetensors = kwargs.pop("use_safetensors", None)
+
+    if use_safetensors is None:
+        use_safetensors = True
+
+    user_agent = {
+        "file_type": "text_inversion",
+        "framework": "pytorch" if from_diffusers else "paddle",
+    }
+    state_dicts = []
+    for pretrained_model_name_or_path in pretrained_model_name_or_paths:
+        if not isinstance(pretrained_model_name_or_path, (dict, paddle.Tensor)):
+            # 3.1. Load textual inversion file
+            model_file = None
+
+            # Let's first try to load .safetensors weights
+            if (use_safetensors and weight_name is None) or (
+                weight_name is not None and weight_name.endswith(".safetensors")
+            ):
+                try:
+                    model_file = _get_model_file(
+                        pretrained_model_name_or_path,
+                        weights_name=weight_name or TEXT_INVERSION_NAME_SAFE,
+                        cache_dir=cache_dir,
+                        force_download=force_download,
+                        resume_download=resume_download,
+                        proxies=proxies,
+                        local_files_only=local_files_only,
+                        use_auth_token=use_auth_token,
+                        revision=revision,
+                        subfolder=subfolder,
+                        user_agent=user_agent,
+                        from_hf_hub=from_hf_hub,
+                        from_aistudio=from_aistudio,
+                    )
+                except Exception:
+                    model_file = None
+                    pass
+
+            if model_file is None:
+                model_file = _get_model_file(
+                    pretrained_model_name_or_path,
+                    weights_name=(weight_name or TORCH_TEXT_INVERSION_NAME)
+                    if from_diffusers
+                    else (weight_name or PADDLE_TEXT_INVERSION_NAME),
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    resume_download=resume_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    use_auth_token=use_auth_token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    user_agent=user_agent,
+                    from_hf_hub=from_hf_hub,
+                    from_aistudio=from_aistudio,
+                )
+            assert model_file is not None, "Could not find the model file!"
+            state_dict = {}
+            load_state_dict(model_file, state_dict)
+        else:
+            state_dict = pretrained_model_name_or_path
+
+        state_dicts.append(state_dict)
+
+    return state_dicts
+
+
+class TextualInversionLoaderMixin:
+    r"""
+    Load Textual Inversion tokens and embeddings to the tokenizer and text encoder.
+    """
+
+    def maybe_convert_prompt(self, prompt: Union[str, List[str]], tokenizer: "PretrainedTokenizer"):  # noqa: F821
+        r"""
+        Processes prompts that include a special token corresponding to a multi-vector textual inversion embedding to
+        be replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
+        inversion token or if the textual inversion token is a single vector, the input prompt is returned.
+
+        Parameters:
+            prompt (`str` or list of `str`):
+                The prompt or prompts to guide the image generation.
+            tokenizer (`PretrainedTokenizer`):
+                The tokenizer responsible for encoding the prompt into input tokens.
+
+        Returns:
+            `str` or list of `str`: The converted prompt
+        """
+        if not isinstance(prompt, List):
+            prompts = [prompt]
+        else:
+            prompts = prompt
+
+        prompts = [self._maybe_convert_prompt(p, tokenizer) for p in prompts]
+
+        if not isinstance(prompt, List):
+            return prompts[0]
+
+        return prompts
+
+    def _maybe_convert_prompt(self, prompt: str, tokenizer: "PretrainedTokenizer"):  # noqa: F821
+        r"""
+        Maybe convert a prompt into a "multi vector"-compatible prompt. If the prompt includes a token that corresponds
+        to a multi-vector textual inversion embedding, this function will process the prompt so that the special token
+        is replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
+        inversion token or a textual inversion token that is a single vector, the input prompt is simply returned.
+
+        Parameters:
+            prompt (`str`):
+                The prompt to guide the image generation.
+            tokenizer (`PretrainedTokenizer`):
+                The tokenizer responsible for encoding the prompt into input tokens.
+
+        Returns:
+            `str`: The converted prompt
+        """
+        tokens = tokenizer.tokenize(prompt)
+        unique_tokens = set(tokens)
+        for token in unique_tokens:
+            if token in tokenizer.added_tokens_encoder:
+                replacement = token
+                i = 1
+                while f"{token}_{i}" in tokenizer.added_tokens_encoder:
+                    replacement += f" {token}_{i}"
+                    i += 1
+
+                prompt = prompt.replace(token, replacement)
+
+        return prompt
+
+    def _check_text_inv_inputs(self, tokenizer, text_encoder, pretrained_model_name_or_paths, tokens):
+        if tokenizer is None:
+            raise ValueError(
+                f"{self.__class__.__name__} requires `self.tokenizer` or passing a `tokenizer` of type `PretrainedTokenizer` for calling"
+                f" `{self.load_textual_inversion.__name__}`"
+            )
+
+        if text_encoder is None:
+            raise ValueError(
+                f"{self.__class__.__name__} requires `self.text_encoder` or passing a `text_encoder` of type `PretrainedModel` for calling"
+                f" `{self.load_textual_inversion.__name__}`"
+            )
+
+        if len(pretrained_model_name_or_paths) > 1 and len(pretrained_model_name_or_paths) != len(tokens):
+            raise ValueError(
+                f"You have passed a list of models of length {len(pretrained_model_name_or_paths)}, and list of tokens of length {len(tokens)} "
+                f"Make sure both lists have the same length."
+            )
+
+        valid_tokens = [t for t in tokens if t is not None]
+        if len(set(valid_tokens)) < len(valid_tokens):
+            raise ValueError(f"You have passed a list of tokens that contains duplicates: {tokens}")
+
+    @staticmethod
+    def _retrieve_tokens_and_embeddings(tokens, state_dicts, tokenizer):
+        all_tokens = []
+        all_embeddings = []
+        for state_dict, token in zip(state_dicts, tokens):
+            if isinstance(state_dict, paddle.Tensor):
+                if token is None:
+                    raise ValueError(
+                        "You are trying to load a textual inversion embedding that has been saved as a PyTorch tensor. Make sure to pass the name of the corresponding token in this case: `token=...`."
+                    )
+                loaded_token = token
+                embedding = state_dict
+            elif len(state_dict) == 1:
+                # diffusers
+                loaded_token, embedding = next(iter(state_dict.items()))
+            elif "string_to_param" in state_dict:
+                # A1111
+                loaded_token = state_dict["name"]
+                embedding = state_dict["string_to_param"]["*"]
+            else:
+                raise ValueError(
+                    f"Loaded state dictonary is incorrect: {state_dict}. \n\n"
+                    "Please verify that the loaded state dictionary of the textual embedding either only has a single key or includes the `string_to_param`"
+                    " input key."
+                )
+
+            if token is not None and loaded_token != token:
+                logger.info(f"The loaded token: {loaded_token} is overwritten by the passed token {token}.")
+            else:
+                token = loaded_token
+
+            if token in tokenizer.get_vocab():
+                raise ValueError(
+                    f"Token {token} already in tokenizer vocabulary. Please choose a different token name or remove {token} and embedding from the tokenizer and text encoder."
+                )
+
+            all_tokens.append(token)
+            all_embeddings.append(embedding)
+
+        return all_tokens, all_embeddings
+
+    @staticmethod
+    def _extend_tokens_and_embeddings(tokens, embeddings, tokenizer):
+        all_tokens = []
+        all_embeddings = []
+
+        for embedding, token in zip(embeddings, tokens):
+            if f"{token}_1" in tokenizer.get_vocab():
+                multi_vector_tokens = [token]
+                i = 1
+                while f"{token}_{i}" in tokenizer.added_tokens_encoder:
+                    multi_vector_tokens.append(f"{token}_{i}")
+                    i += 1
+
+                raise ValueError(
+                    f"Multi-vector Token {multi_vector_tokens} already in tokenizer vocabulary. Please choose a different token name or remove the {multi_vector_tokens} and embedding from the tokenizer and text encoder."
+                )
+
+            is_multi_vector = len(embedding.shape) > 1 and embedding.shape[0] > 1
+            if is_multi_vector:
+                all_tokens += [token] + [f"{token}_{i}" for i in range(1, embedding.shape[0])]
+                all_embeddings += [e for e in embedding]  # noqa: C416
+            else:
+                all_tokens += [token]
+                all_embeddings += [embedding[0]] if len(embedding.shape) > 1 else [embedding]
+
+        return all_tokens, all_embeddings
+
+    def load_textual_inversion(
+        self,
+        pretrained_model_name_or_path: Union[str, List[str], Dict[str, paddle.Tensor], List[Dict[str, paddle.Tensor]]],
+        token: Optional[Union[str, List[str]]] = None,
+        tokenizer: Optional["PretrainedTokenizer"] = None,  # noqa: F821
+        text_encoder: Optional["PretrainedModel"] = None,  # noqa: F821
+        **kwargs,
+    ):
+        r"""
+        Load Textual Inversion embeddings into the text encoder of [`StableDiffusionPipeline`] (both 🤗 Diffusers and
+        Automatic1111 formats are supported).
+
+        Parameters:
+            pretrained_model_name_or_path (`str` or `os.PathLike` or `List[str or os.PathLike]` or `Dict` or `List[Dict]`):
+                Can be either one of the following or a list of them:
+
+                    - A string, the *model id* (for example `sd-concepts-library/low-poly-hd-logos-icons`) of a
+                      pretrained model hosted on the Hub.
+                    - A path to a *directory* (for example `./my_text_inversion_directory/`) containing the textual
+                      inversion weights.
+                    - A path to a *file* (for example `./my_text_inversions.pt`) containing textual inversion weights.
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+            token (`str` or `List[str]`, *optional*):
+                Override the token to use for the textual inversion weights. If `pretrained_model_name_or_path` is a
+                list, then `token` must also be a list of equal length.
+            text_encoder ([`~transformers.CLIPTextModel`], *optional*):
+                Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+                If not specified, function will take self.tokenizer.
+            tokenizer ([`~transformers.CLIPTokenizer`], *optional*):
+                A `CLIPTokenizer` to tokenize text. If not specified, function will take self.tokenizer.
+            weight_name (`str`, *optional*):
+                Name of a custom weight file. This should be used when:
+
+                    - The saved textual inversion file is in 🤗 Diffusers format, but was saved under a specific weight
+                      name such as `text_inv.bin`.
+                    - The saved textual inversion file is in the Automatic1111 format.
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            use_auth_token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+            mirror (`str`, *optional*):
+                Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
+                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+                information.
+
+        Example:
+
+        To load a Textual Inversion embedding vector in 🤗 Diffusers format:
+
+        ```py
+        from ppdiffusers import StableDiffusionPipeline
+        import paddle
+
+        model_id = "runwayml/stable-diffusion-v1-5"
+        pipe = StableDiffusionPipeline.from_pretrained(model_id, paddle_dtype=paddle.float16)
+
+        pipe.load_textual_inversion("sd-concepts-library/cat-toy")
+
+        prompt = "A <cat-toy> backpack"
+
+        image = pipe(prompt, num_inference_steps=50).images[0]
+        image.save("cat-backpack.png")
+        ```
+
+        To load a Textual Inversion embedding vector in Automatic1111 format, make sure to download the vector first
+        (for example from [civitAI](https://civitai.com/models/3036?modelVersionId=9857)) and then load the vector
+        locally:
+
+        ```py
+        from ppdiffusers import StableDiffusionPipeline
+        import paddle
+
+        model_id = "runwayml/stable-diffusion-v1-5"
+        pipe = StableDiffusionPipeline.from_pretrained(model_id, paddle_dtype=paddle.float16)
+
+        pipe.load_textual_inversion("./charturnerv2.pt", token="charturnerv2")
+
+        prompt = "charturnerv2, multiple views of the same character in the same outfit, a character turnaround of a woman wearing a black jacket and red shirt, best quality, intricate details."
+
+        image = pipe(prompt, num_inference_steps=50).images[0]
+        image.save("character.png")
+        ```
+
+        """
+        # 1. Set correct tokenizer and text encoder
+        tokenizer = tokenizer or getattr(self, "tokenizer", None)
+        text_encoder = text_encoder or getattr(self, "text_encoder", None)
+
+        # 2. Normalize inputs
+        pretrained_model_name_or_paths = (
+            [pretrained_model_name_or_path]
+            if not isinstance(pretrained_model_name_or_path, list)
+            else pretrained_model_name_or_path
+        )
+        tokens = [token] if not isinstance(token, list) else token
+        if tokens[0] is None:
+            tokens = tokens * len(pretrained_model_name_or_paths)
+
+        # 3. Check inputs
+        self._check_text_inv_inputs(tokenizer, text_encoder, pretrained_model_name_or_paths, tokens)
+
+        # 4. Load state dicts of textual embeddings
+        state_dicts = load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
+
+        # 4.1 Handle the special case when state_dict is a tensor that contains n embeddings for n tokens
+        if len(tokens) > 1 and len(state_dicts) == 1:
+            if isinstance(state_dicts[0], paddle.Tensor):
+                state_dicts = list(state_dicts[0])
+                if len(tokens) != len(state_dicts):
+                    raise ValueError(
+                        f"You have passed a state_dict contains {len(state_dicts)} embeddings, and list of tokens of length {len(tokens)} "
+                        f"Make sure both have the same length."
+                    )
+
+        # 4. Retrieve tokens and embeddings
+        tokens, embeddings = self._retrieve_tokens_and_embeddings(tokens, state_dicts, tokenizer)
+
+        # 5. Extend tokens and embeddings for multi vector
+        tokens, embeddings = self._extend_tokens_and_embeddings(tokens, embeddings, tokenizer)
+
+        # 6. Make sure all embeddings have the correct size
+        expected_emb_dim = text_encoder.get_input_embeddings().weight.shape[-1]
+        if any(expected_emb_dim != emb.shape[-1] for emb in embeddings):
+            raise ValueError(
+                "Loaded embeddings are of incorrect shape. Expected each textual inversion embedding "
+                f"to be of shape {expected_emb_dim}, but are {expected_emb_dim} "
+            )
+
+        # 7.2 save expected device and dtype
+        dtype = text_encoder.dtype
+
+        # 7.3 Increase token embedding matrix
+        text_encoder.resize_token_embeddings(len(tokenizer) + len(tokens))
+
+        # 7.4 Load token and embedding
+        with paddle.no_grad():
+            for token, embedding in zip(tokens, embeddings):
+                # add tokens and get ids
+                tokenizer.add_tokens(token)
+                token_id = tokenizer.convert_tokens_to_ids(token)
+                text_encoder.get_input_embeddings().weight[token_id] = embedding
+                logger.info(f"Loaded textual inversion embedding for {token}.")
+
+        text_encoder.get_input_embeddings().to(
+            dtype=dtype,
+        )

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/loaders/unet.py

@@ -0,0 +1,830 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+from collections import defaultdict
+from contextlib import nullcontext
+from functools import partial
+from typing import Callable, Dict, List, Optional, Union
+
+import paddle
+
+from ..models.embeddings import ImageProjection
+from ..models.modeling_pytorch_paddle_utils import (
+    convert_paddle_state_dict_to_pytorch,
+    convert_pytorch_state_dict_to_paddle,
+)
+from ..models.modeling_utils import faster_set_state_dict, load_state_dict
+from ..utils import (
+    DIFFUSERS_CACHE,
+    FROM_AISTUDIO,
+    FROM_DIFFUSERS,
+    FROM_HF_HUB,
+    HF_HUB_OFFLINE,
+    LOW_CPU_MEM_USAGE_DEFAULT,
+    PPDIFFUSERS_CACHE,
+    TO_DIFFUSERS,
+    USE_PEFT_BACKEND,
+    _get_model_file,
+    delete_adapter_layers,
+    is_paddle_version,
+    is_ppxformers_available,
+    is_safetensors_available,
+    is_torch_available,
+    logging,
+    set_adapter_layers,
+    set_weights_and_activate_adapters,
+)
+from .utils import AttnProcsLayers
+
+if is_safetensors_available():
+    from safetensors.numpy import save_file as np_safe_save_file
+
+    if is_torch_available():
+        from safetensors.torch import save_file as torch_safe_save_file
+
+logger = logging.get_logger(__name__)
+
+
+if is_torch_available():
+    import torch
+TEXT_ENCODER_NAME = "text_encoder"
+UNET_NAME = "unet"
+
+TORCH_LORA_WEIGHT_NAME = "pytorch_lora_weights.bin"
+TORCH_LORA_WEIGHT_NAME_SAFE = "pytorch_lora_weights.safetensors"
+
+PADDLE_LORA_WEIGHT_NAME = "paddle_lora_weights.pdparams"
+PADDLE_LORA_WEIGHT_NAME_SAFE = "paddle_lora_weights.safetensors"
+
+TORCH_CUSTOM_DIFFUSION_WEIGHT_NAME = "pytorch_custom_diffusion_weights.bin"
+TORCH_CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE = "pytorch_custom_diffusion_weights.safetensors"
+
+PADDLE_CUSTOM_DIFFUSION_WEIGHT_NAME = "paddle_custom_diffusion_weights.pdparams"
+PADDLE_CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE = "paddle_custom_diffusion_weights.safetensors"
+
+
+class UNet2DConditionLoadersMixin:
+    """
+    Load LoRA layers into a [`UNet2DCondtionModel`].
+    """
+
+    text_encoder_name = TEXT_ENCODER_NAME
+    unet_name = UNET_NAME
+
+    def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, paddle.Tensor]], **kwargs):
+        r"""
+        Load pretrained attention processor layers into [`UNet2DConditionModel`]. Attention processor layers have to be
+        defined in
+        [`attention_processor.py`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py)
+        and be a `nn.Layer` class.
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                Can be either:
+
+                    - A string, the model id (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a directory (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            use_auth_token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+                argument to `True` will raise an error.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+            mirror (`str`, *optional*):
+                Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
+                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+                information.
+
+        Example:
+
+        ```py
+        from ppdiffusers import AutoPipelineForText2Image
+        import torch
+
+        pipeline = AutoPipelineForText2Image.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", paddle_dtype=paddle.float16
+        )
+        pipeline.unet.load_attn_procs(
+            "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+        )
+        ```
+        """
+        from ..models.attention_processor import CustomDiffusionAttnProcessor
+        from ..models.lora import (
+            LoRACompatibleConv,
+            LoRACompatibleLinear,
+            LoRAConv2dLayer,
+            LoRALinearLayer,
+        )
+
+        from_hf_hub = kwargs.pop("from_hf_hub", FROM_HF_HUB)
+        from_aistudio = kwargs.pop("from_aistudio", FROM_AISTUDIO)
+        cache_dir = kwargs.pop("cache_dir", None)
+        if cache_dir is None:
+            if from_aistudio:
+                cache_dir = None  # TODO, check aistudio cache
+            elif from_hf_hub:
+                cache_dir = DIFFUSERS_CACHE
+            else:
+                cache_dir = PPDIFFUSERS_CACHE
+        from_diffusers = kwargs.pop("from_diffusers", FROM_DIFFUSERS)
+
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
+        revision = kwargs.pop("revision", None)
+        subfolder = kwargs.pop("subfolder", None)
+        if subfolder is None:
+            subfolder = ""
+        weight_name = kwargs.pop("weight_name", None)
+        use_safetensors = kwargs.pop("use_safetensors", None)
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", LOW_CPU_MEM_USAGE_DEFAULT)
+        # This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
+        # See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
+        network_alphas = kwargs.pop("network_alphas", None)
+
+        _pipeline = kwargs.pop("_pipeline", None)  # noqa: F841
+
+        is_network_alphas_none = network_alphas is None
+
+        if use_safetensors is None:
+            use_safetensors = True
+
+        if weight_name is not None:
+            if "paddle" in weight_name.lower() or ".pdparams" in weight_name.lower():
+                if from_diffusers:
+                    logger.warning(
+                        "Detect the weight is in ppdiffusers format, but currently, `from_diffusers` is set to `True`. To proceed, we will change the value of `from_diffusers` to `False`!"
+                    )
+                    from_diffusers = False
+            elif "torch" in weight_name.lower() or ".bin" in weight_name.lower() or ".pt" in weight_name.lower():
+                if not from_diffusers:
+                    logger.warning(
+                        "Detect the weight is in diffusers format, but currently, `from_diffusers` is set to `False`. To proceed, we will change the value of `from_diffusers` to `True`!"
+                    )
+                    from_diffusers = True
+
+        user_agent = {
+            "file_type": "attn_procs_weights",
+            "framework": "pytorch" if from_diffusers else "paddle",
+        }
+
+        if low_cpu_mem_usage and (not is_paddle_version(">=", "2.5.0") and not is_paddle_version("==", "0.0.0")):
+            raise NotImplementedError(
+                "Low memory initialization requires paddlepaddle-gpu >= 2.5.0. Please either update your PaddlePaddle version or set"
+                " `low_cpu_mem_usage=False`."
+            )
+
+        model_file = None
+        state_dict = {}
+        if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+            # Let's first try to load .safetensors weights
+            if (use_safetensors and weight_name is None) or (
+                weight_name is not None and weight_name.endswith(".safetensors")
+            ):
+                try:
+                    model_file = _get_model_file(
+                        pretrained_model_name_or_path_or_dict,
+                        weights_name=(weight_name or TORCH_LORA_WEIGHT_NAME_SAFE)
+                        if from_diffusers
+                        else ((weight_name or PADDLE_LORA_WEIGHT_NAME_SAFE)),
+                        cache_dir=cache_dir,
+                        force_download=force_download,
+                        resume_download=resume_download,
+                        proxies=proxies,
+                        local_files_only=local_files_only,
+                        use_auth_token=use_auth_token,
+                        revision=revision,
+                        subfolder=subfolder,
+                        user_agent=user_agent,
+                        from_aistudio=from_aistudio,
+                        from_hf_hub=from_hf_hub,
+                    )
+                except Exception:
+                    model_file = None
+            if model_file is None:
+                model_file = _get_model_file(
+                    pretrained_model_name_or_path_or_dict,
+                    weights_name=(weight_name or TORCH_LORA_WEIGHT_NAME)
+                    if from_diffusers
+                    else ((weight_name or PADDLE_LORA_WEIGHT_NAME)),
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    resume_download=resume_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    use_auth_token=use_auth_token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    user_agent=user_agent,
+                    from_aistudio=from_aistudio,
+                    from_hf_hub=from_hf_hub,
+                )
+
+            assert model_file is not None, "Could not find the model file!"
+            data_format = load_state_dict(model_file, state_dict)
+            if not from_diffusers and data_format == "pt":
+                logger.warning(
+                    "Detect the weight is in diffusers format, but currently, `from_diffusers` is set to `False`. To proceed, we will change the value of `from_diffusers` to `True`!"
+                )
+                from_diffusers = True
+            if from_diffusers and data_format in ["pd", "np"]:
+                logger.warning(
+                    "Detect the weight is in ppdiffusers format, but currently, `from_diffusers` is set to `True`. To proceed, we will change the value of `from_diffusers` to `False`!"
+                )
+                from_diffusers = False
+        else:
+            state_dict = pretrained_model_name_or_path_or_dict
+
+        # fill attn processors
+        lora_layers_list = []
+
+        is_lora = all(("lora" in k or k.endswith(".alpha")) for k in state_dict.keys()) and not USE_PEFT_BACKEND
+        is_custom_diffusion = any("custom_diffusion" in k for k in state_dict.keys())
+
+        if is_lora:
+            # correct keys
+            state_dict, network_alphas = self.convert_state_dict_legacy_attn_format(state_dict, network_alphas)
+
+            if network_alphas is not None:
+                network_alphas_keys = list(network_alphas.keys())
+                used_network_alphas_keys = set()
+
+            lora_grouped_dict = defaultdict(dict)
+            mapped_network_alphas = {}
+
+            all_keys = list(state_dict.keys())
+            for key in all_keys:
+                value = state_dict.pop(key)
+                attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
+                lora_grouped_dict[attn_processor_key][sub_key] = value
+
+                # Create another `mapped_network_alphas` dictionary so that we can properly map them.
+                if network_alphas is not None:
+                    for k in network_alphas_keys:
+                        if k.replace(".alpha", "") in key:
+                            mapped_network_alphas.update({attn_processor_key: network_alphas.get(k)})
+                            used_network_alphas_keys.add(k)
+
+            if not is_network_alphas_none:
+                if len(set(network_alphas_keys) - used_network_alphas_keys) > 0:
+                    raise ValueError(
+                        f"The `network_alphas` has to be empty at this point but has the following keys \n\n {', '.join(network_alphas.keys())}"
+                    )
+
+            if len(state_dict) > 0:
+                raise ValueError(
+                    f"The `state_dict` has to be empty at this point but has the following keys \n\n {', '.join(state_dict.keys())}"
+                )
+
+            for key, value_dict in lora_grouped_dict.items():
+                attn_processor = self
+                for sub_key in key.split("."):
+                    attn_processor = getattr(attn_processor, sub_key)
+
+                # Process non-attention layers, which don't have to_{k,v,q,out_proj}_lora layers
+                # or add_{k,v,q,out_proj}_proj_lora layers.
+                if from_diffusers:
+                    rank = value_dict["lora.down.weight"].shape[0]
+                else:
+                    rank = value_dict["lora.down.weight"].shape[1]
+
+                if isinstance(attn_processor, LoRACompatibleConv):
+                    in_features = attn_processor._in_channels
+                    out_features = attn_processor._out_channels
+                    kernel_size = attn_processor._kernel_size
+
+                    ctx = paddle.LazyGuard if low_cpu_mem_usage else nullcontext
+                    with ctx():
+                        lora = LoRAConv2dLayer(
+                            in_features=in_features,
+                            out_features=out_features,
+                            rank=rank,
+                            kernel_size=kernel_size,
+                            stride=attn_processor._stride,
+                            padding=attn_processor._padding,
+                            network_alpha=mapped_network_alphas.get(key),
+                        )
+                elif isinstance(attn_processor, LoRACompatibleLinear):
+                    ctx = paddle.LazyGuard if low_cpu_mem_usage else nullcontext
+                    with ctx():
+                        lora = LoRALinearLayer(
+                            attn_processor.in_features,
+                            attn_processor.out_features,
+                            rank,
+                            mapped_network_alphas.get(key),
+                        )
+                else:
+                    raise ValueError(f"Module {key} is not a LoRACompatibleConv or LoRACompatibleLinear module.")
+
+                value_dict = {k.replace("lora.", ""): v for k, v in value_dict.items()}
+                lora_layers_list.append((attn_processor, lora))
+                if from_diffusers:
+                    convert_pytorch_state_dict_to_paddle(lora, value_dict)
+                faster_set_state_dict(lora, value_dict)
+
+        elif is_custom_diffusion:
+            attn_processors = {}
+            custom_diffusion_grouped_dict = defaultdict(dict)
+            for key, value in state_dict.items():
+                if len(value) == 0:
+                    custom_diffusion_grouped_dict[key] = {}
+                else:
+                    if "to_out" in key:
+                        attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
+                    else:
+                        attn_processor_key, sub_key = ".".join(key.split(".")[:-2]), ".".join(key.split(".")[-2:])
+                    custom_diffusion_grouped_dict[attn_processor_key][sub_key] = value
+
+            for key, value_dict in custom_diffusion_grouped_dict.items():
+                if len(value_dict) == 0:
+                    attn_processors[key] = CustomDiffusionAttnProcessor(
+                        train_kv=False, train_q_out=False, hidden_size=None, cross_attention_dim=None
+                    )
+                else:
+                    if from_diffusers:
+                        cross_attention_dim = value_dict["to_k_custom_diffusion.weight"].shape[1]
+                        hidden_size = value_dict["to_k_custom_diffusion.weight"].shape[0]
+                    else:
+                        cross_attention_dim = value_dict["to_k_custom_diffusion.weight"].shape[0]
+                        hidden_size = value_dict["to_k_custom_diffusion.weight"].shape[1]
+                    train_q_out = True if "to_q_custom_diffusion.weight" in value_dict else False
+                    attn_processors[key] = CustomDiffusionAttnProcessor(
+                        train_kv=True,
+                        train_q_out=train_q_out,
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                    )
+                    if from_diffusers:
+                        convert_pytorch_state_dict_to_paddle(attn_processors[key], value_dict)
+                    faster_set_state_dict(attn_processors[key], value_dict)
+        elif USE_PEFT_BACKEND:
+            # In that case we have nothing to do as loading the adapter weights is already handled above by `set_peft_model_state_dict`
+            # on the Unet
+            pass
+        else:
+            raise ValueError(
+                f"{model_file} does not seem to be in the correct format expected by LoRA or Custom Diffusion training."
+            )
+
+        # For PEFT backend the Unet is already offloaded at this stage as it is handled inside `lora_lora_weights_into_unet`
+        if not USE_PEFT_BACKEND:
+
+            # only custom diffusion needs to set attn processors
+            if is_custom_diffusion:
+                self.set_attn_processor(attn_processors)
+
+            # set lora layers
+            for target_module, lora_layer in lora_layers_list:
+                target_module.set_lora_layer(lora_layer)
+
+            self.to(dtype=self.dtype)
+
+    def convert_state_dict_legacy_attn_format(self, state_dict, network_alphas):
+        is_new_lora_format = all(
+            key.startswith(self.unet_name) or key.startswith(self.text_encoder_name) for key in state_dict.keys()
+        )
+        if is_new_lora_format:
+            # Strip the `"unet"` prefix.
+            is_text_encoder_present = any(key.startswith(self.text_encoder_name) for key in state_dict.keys())
+            if is_text_encoder_present:
+                warn_message = "The state_dict contains LoRA params corresponding to the text encoder which are not being used here. To use both UNet and text encoder related LoRA params, use [`pipe.load_lora_weights()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.load_lora_weights)."
+                logger.warn(warn_message)
+            unet_keys = [k for k in state_dict.keys() if k.startswith(self.unet_name)]
+            state_dict = {k.replace(f"{self.unet_name}.", ""): v for k, v in state_dict.items() if k in unet_keys}
+
+        # change processor format to 'pure' LoRACompatibleLinear format
+        if any("processor" in k.split(".") for k in state_dict.keys()):
+
+            def format_to_lora_compatible(key):
+                if "processor" not in key.split("."):
+                    return key
+                return key.replace(".processor", "").replace("to_out_lora", "to_out.0.lora").replace("_lora", ".lora")
+
+            state_dict = {format_to_lora_compatible(k): v for k, v in state_dict.items()}
+
+            if network_alphas is not None:
+                network_alphas = {format_to_lora_compatible(k): v for k, v in network_alphas.items()}
+        return state_dict, network_alphas
+
+    def save_attn_procs(
+        self,
+        save_directory: Union[str, os.PathLike],
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+        to_diffusers: Optional[bool] = None,
+        **kwargs,
+    ):
+        r"""
+        Save attention processor layers to a directory so that it can be reloaded with the
+        [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`] method.
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to save an attention processor to (will be created if it doesn't exist).
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful during distributed training and you
+                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+                process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful during distributed training when you need to
+                replace `torch.save` with another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or with `pickle`.
+
+        Example:
+
+        ```py
+        import paddle
+        from ppdiffusers import DiffusionPipeline
+
+        pipeline = DiffusionPipeline.from_pretrained(
+            "CompVis/stable-diffusion-v1-4",
+            paddle_dtype=paddle.float16,
+        )
+        pipeline.unet.load_attn_procs("path-to-save-model", weight_name="paddle_custom_diffusion_weights.pdparams")
+        pipeline.unet.save_attn_procs("path-to-save-model", weight_name="paddle_custom_diffusion_weights.pdparams")
+        ```
+        """
+        from ..models.attention_processor import (
+            CustomDiffusionAttnProcessor,
+            CustomDiffusionAttnProcessor2_5,
+            CustomDiffusionXFormersAttnProcessor,
+        )
+
+        if to_diffusers is None:
+            to_diffusers = TO_DIFFUSERS
+        if os.path.isfile(save_directory):
+            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
+            return
+
+        os.makedirs(save_directory, exist_ok=True)
+
+        is_custom_diffusion = any(
+            isinstance(
+                x,
+                (CustomDiffusionAttnProcessor, CustomDiffusionAttnProcessor2_5, CustomDiffusionXFormersAttnProcessor),
+            )
+            for (_, x) in self.attn_processors.items()
+        )
+        if is_custom_diffusion:
+            model_to_save = AttnProcsLayers(
+                {
+                    y: x
+                    for (y, x) in self.attn_processors.items()
+                    if isinstance(
+                        x,
+                        (
+                            CustomDiffusionAttnProcessor,
+                            CustomDiffusionAttnProcessor2_5,
+                            CustomDiffusionXFormersAttnProcessor,
+                        ),
+                    )
+                }
+            )
+            state_dict = model_to_save.state_dict()
+            for name, attn in self.attn_processors.items():
+                if len(attn.state_dict()) == 0:
+                    state_dict[name] = {}
+        else:
+            model_to_save = AttnProcsLayers(self.attn_processors)
+            state_dict = model_to_save.state_dict()
+
+        if weight_name is None:
+            if to_diffusers:
+                if safe_serialization:
+                    weight_name = (
+                        TORCH_CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE if is_custom_diffusion else TORCH_LORA_WEIGHT_NAME_SAFE
+                    )
+                else:
+                    weight_name = TORCH_CUSTOM_DIFFUSION_WEIGHT_NAME if is_custom_diffusion else TORCH_LORA_WEIGHT_NAME
+            else:
+                if safe_serialization:
+                    weight_name = (
+                        PADDLE_CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE
+                        if is_custom_diffusion
+                        else PADDLE_LORA_WEIGHT_NAME_SAFE
+                    )
+                else:
+                    weight_name = (
+                        PADDLE_CUSTOM_DIFFUSION_WEIGHT_NAME if is_custom_diffusion else PADDLE_LORA_WEIGHT_NAME
+                    )
+        else:
+            if "paddle" in weight_name.lower() or "pdparams" in weight_name.lower():
+                to_diffusers = False
+            elif "torch" in weight_name.lower() or "bin" in weight_name.lower():
+                to_diffusers = True
+
+        # choose save_function
+        if save_function is None:
+            if to_diffusers:
+                if not is_torch_available() and not safe_serialization:
+                    safe_serialization = True
+                    logger.warning(
+                        "PyTorch is not installed, and `safe_serialization` is currently set to `False`. "
+                        "To ensure proper model saving, we will automatically set `safe_serialization=True`. "
+                        "If you want to keep `safe_serialization=False`, please make sure PyTorch is installed."
+                    )
+                if safe_serialization:
+                    if is_torch_available():
+                        save_function = partial(torch_safe_save_file, metadata={"format": "pt"})
+                    else:
+                        save_function = partial(np_safe_save_file, metadata={"format": "pt"})
+                else:
+                    save_function = torch.save
+
+                convert_paddle_state_dict_to_pytorch(self, state_dict)
+            else:
+                if safe_serialization:
+                    for k, v in state_dict.items():
+                        if isinstance(v, paddle.Tensor):
+                            state_dict[k] = v.cpu().numpy()
+
+                    save_function = partial(np_safe_save_file, metadata={"format": "pd"})
+                else:
+                    save_function = paddle.save
+
+        # Save the model
+        save_function(state_dict, os.path.join(save_directory, weight_name))
+        logger.info(f"Model weights saved in {os.path.join(save_directory, weight_name)}")
+
+    def fuse_lora(self, lora_scale=1.0, safe_fusing=False):
+        self.lora_scale = lora_scale
+        self._safe_fusing = safe_fusing
+        self.apply(self._fuse_lora_apply)
+
+    def _fuse_lora_apply(self, module):
+        if not USE_PEFT_BACKEND:
+            if hasattr(module, "_fuse_lora"):
+                module._fuse_lora(self.lora_scale, self._safe_fusing)
+        else:
+            from ppdiffusers.peft.tuners.tuners_utils import BaseTunerLayer
+
+            if isinstance(module, BaseTunerLayer):
+                if self.lora_scale != 1.0:
+                    module.scale_layer(self.lora_scale)
+                module.merge(safe_merge=self._safe_fusing)
+
+    def unfuse_lora(self):
+        self.apply(self._unfuse_lora_apply)
+
+    def _unfuse_lora_apply(self, module):
+        if not USE_PEFT_BACKEND:
+            if hasattr(module, "_unfuse_lora"):
+                module._unfuse_lora()
+        else:
+            from ppdiffusers.peft.tuners.tuners_utils import BaseTunerLayer
+
+            if isinstance(module, BaseTunerLayer):
+                module.unmerge()
+
+    def set_adapters(
+        self,
+        adapter_names: Union[List[str], str],
+        weights: Optional[Union[List[float], float]] = None,
+    ):
+        """
+        Set the currently active adapters for use in the UNet.
+
+        Args:
+            adapter_names (`List[str]` or `str`):
+                The names of the adapters to use.
+            adapter_weights (`Union[List[float], float]`, *optional*):
+                The adapter(s) weights to use with the UNet. If `None`, the weights are set to `1.0` for all the
+                adapters.
+
+        Example:
+
+        ```py
+        from ppdiffusers import AutoPipelineForText2Image
+        import torch
+
+        pipeline = AutoPipelineForText2Image.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", paddle_dtype=paddle.float16
+        )
+        pipeline.load_lora_weights(
+            "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+        )
+        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
+        pipeline.set_adapters(["cinematic", "pixel"], adapter_weights=[0.5, 0.5])
+        ```
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for `set_adapters()`.")
+
+        adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
+
+        if weights is None:
+            weights = [1.0] * len(adapter_names)
+        elif isinstance(weights, float):
+            weights = [weights] * len(adapter_names)
+
+        if len(adapter_names) != len(weights):
+            raise ValueError(
+                f"Length of adapter names {len(adapter_names)} is not equal to the length of their weights {len(weights)}."
+            )
+
+        set_weights_and_activate_adapters(self, adapter_names, weights)
+
+    def disable_lora(self):
+        """
+        Disable the UNet's active LoRA layers.
+
+        Example:
+
+        ```py
+        from ppdiffusers import AutoPipelineForText2Image
+        import torch
+
+        pipeline = AutoPipelineForText2Image.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", paddle_dtype=paddle.float16
+        )
+        pipeline.load_lora_weights(
+            "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+        )
+        pipeline.disable_lora()
+        ```
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+        set_adapter_layers(self, enabled=False)
+
+    def enable_lora(self):
+        """
+        Enable the UNet's active LoRA layers.
+
+        Example:
+
+        ```py
+        from ppdiffusers import AutoPipelineForText2Image
+        import paddle
+
+        pipeline = AutoPipelineForText2Image.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", paddle_dtype=paddle.float16
+        )
+        pipeline.load_lora_weights(
+            "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+        )
+        pipeline.enable_lora()
+        ```
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+        set_adapter_layers(self, enabled=True)
+
+    def delete_adapters(self, adapter_names: Union[List[str], str]):
+        """
+        Delete an adapter's LoRA layers from the UNet.
+
+        Args:
+            adapter_names (`Union[List[str], str]`):
+                The names (single string or list of strings) of the adapter to delete.
+
+        Example:
+
+        ```py
+        from ppdiffusers import AutoPipelineForText2Image
+        import paddle
+
+        pipeline = AutoPipelineForText2Image.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", paddle_dtype=paddle.float16
+        )
+        pipeline.load_lora_weights(
+            "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_names="cinematic"
+        )
+        pipeline.delete_adapters("cinematic")
+        ```
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        if isinstance(adapter_names, str):
+            adapter_names = [adapter_names]
+
+        for adapter_name in adapter_names:
+            delete_adapter_layers(self, adapter_name)
+
+            # Pop also the corresponding adapter from the config
+            if hasattr(self, "peft_config"):
+                self.peft_config.pop(adapter_name, None)
+
+    def _load_ip_adapter_weights(self, state_dict, from_diffusers=None):
+        if from_diffusers is None:
+            from_diffusers = FROM_DIFFUSERS
+
+        str_dtype = str(self.dtype).replace("paddle.", "")
+        from ..models.attention_processor import (
+            AttnProcessor,
+            AttnProcessor2_5,
+            IPAdapterAttnProcessor,
+            IPAdapterAttnProcessor2_5,
+        )
+
+        # set ip-adapter cross-attention processors & load state_dict
+        attn_procs = {}
+        key_id = 1
+        for name in self.attn_processors.keys():
+            cross_attention_dim = None if name.endswith("attn1.processor") else self.config.cross_attention_dim
+            if name.startswith("mid_block"):
+                hidden_size = self.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(self.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = self.config.block_out_channels[block_id]
+            if cross_attention_dim is None or "motion_modules" in name:
+                attn_processor_class = AttnProcessor2_5 if is_ppxformers_available() else AttnProcessor
+                attn_procs[name] = attn_processor_class()
+            else:
+                attn_processor_class = (
+                    IPAdapterAttnProcessor2_5 if is_ppxformers_available() else IPAdapterAttnProcessor
+                )
+                attn_procs[name] = attn_processor_class(
+                    hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, scale=1.0
+                ).to(dtype=self.dtype)
+
+                value_dict = {}
+                for k, w in attn_procs[name].state_dict().items():
+                    value_dict.update({f"{k}": state_dict["ip_adapter"][f"{key_id}.{k}"].astype(str_dtype)})
+
+                if from_diffusers:
+                    convert_pytorch_state_dict_to_paddle(attn_procs[name], value_dict)
+                attn_procs[name].load_dict(value_dict)
+                key_id += 2
+
+        self.set_attn_processor(attn_procs)
+
+        # create image projection layers.
+        if from_diffusers:
+            clip_embeddings_dim = state_dict["image_proj"]["proj.weight"].shape[-1]
+            cross_attention_dim = state_dict["image_proj"]["proj.weight"].shape[0] // 4
+        else:
+            clip_embeddings_dim = state_dict["image_proj"]["proj.weight"].shape[0]
+            cross_attention_dim = state_dict["image_proj"]["proj.weight"].shape[-1] // 4
+        image_projection = ImageProjection(
+            cross_attention_dim=cross_attention_dim, image_embed_dim=clip_embeddings_dim, num_image_text_embeds=4
+        )
+        image_projection.to(dtype=self.dtype)
+
+        # load image projection layer weights
+        image_proj_state_dict = {}
+        image_proj_state_dict.update(
+            {
+                "image_embeds.weight": state_dict["image_proj"]["proj.weight"].astype(str_dtype),
+                "image_embeds.bias": state_dict["image_proj"]["proj.bias"].astype(str_dtype),
+                "norm.weight": state_dict["image_proj"]["norm.weight"].astype(str_dtype),
+                "norm.bias": state_dict["image_proj"]["norm.bias"].astype(str_dtype),
+            }
+        )
+        if from_diffusers:
+            convert_pytorch_state_dict_to_paddle(image_projection, image_proj_state_dict)
+        image_projection.load_dict(image_proj_state_dict)
+
+        self.encoder_hid_proj = image_projection.to(dtype=self.dtype)
+        self.config.encoder_hid_dim_type = "ip_image_proj"
+
+    delete_adapter_layers

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/loaders/utils.py

@@ -0,0 +1,121 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import functools
+import weakref
+from collections import OrderedDict
+
+import paddle
+import paddle.nn as nn
+
+try:
+    from paddle.nn.layer.layers import HookRemoveHelper
+except ImportError:
+    from paddle.fluid.dygraph.layers import HookRemoveHelper
+
+from typing import Any, Callable, Dict, Optional
+
+
+class _WrappedHook:
+    def __init__(self, hook: Callable, module: Optional["nn.Layer"] = None):
+        self.hook: Callable = hook
+        functools.update_wrapper(self, hook)
+
+        self.with_module: bool = False
+
+        if module is not None:
+            self.module: weakref.ReferenceType["nn.Layer"] = weakref.ref(module)
+            self.with_module = True
+
+    def __call__(self, *args: Any, **kwargs: Any) -> Any:
+        if self.with_module:
+            module = self.module()
+            if module is None:
+                raise RuntimeError("You are trying to call the hook of a dead Module!")
+            return self.hook(module, *args, **kwargs)
+        return self.hook(*args, **kwargs)
+
+    def __getstate__(self) -> Dict:
+        result = {"hook": self.hook, "with_module": self.with_module}
+        if self.with_module:
+            result["module"] = self.module()
+
+        return result
+
+    def __setstate__(self, state: Dict):
+        self.hook = state["hook"]
+        self.with_module = state["with_module"]
+
+        if self.with_module:
+            if state["module"] is None:
+                raise RuntimeError("You are trying to revive the hook of a dead Module!")
+            self.module = weakref.ref(state["module"])
+
+
+class AttnProcsLayers(nn.Layer):
+    def __init__(self, state_dict: Dict[str, paddle.Tensor]):
+        super().__init__()
+        self.layers = nn.LayerList(state_dict.values())
+        self.mapping = dict(enumerate(state_dict.keys()))
+        self.rev_mapping = {v: k for k, v in enumerate(state_dict.keys())}
+
+        # .processor for unet, .self_attn for text encoder
+        self.split_keys = [".processor", ".self_attn"]
+
+        # we add a hook to state_dict() and load_state_dict() so that the
+        # naming fits with `unet.attn_processors`
+        def map_to(state_dict, *args, **kwargs):
+            new_state_dict = {}
+            for key, value in state_dict.items():
+                num = int(key.split(".")[1])  # 0 is always "layers"
+                new_key = key.replace(f"layers.{num}", self.mapping[num])
+                new_state_dict[new_key] = value
+
+            return new_state_dict
+
+        def remap_key(key, state_dict):
+            for k in self.split_keys:
+                if k in key:
+                    return key.split(k)[0] + k
+
+            raise ValueError(
+                f"There seems to be a problem with the state_dict: {set(state_dict.keys())}. {key} has to have one of {self.split_keys}."
+            )
+
+        def map_from(module, state_dict, *args, **kwargs):
+            all_keys = list(state_dict.keys())
+            for key in all_keys:
+                replace_key = remap_key(key, state_dict)
+                new_key = key.replace(replace_key, f"layers.{module.rev_mapping[replace_key]}")
+                state_dict[new_key] = state_dict[key]
+                del state_dict[key]
+
+        self.register_state_dict_hook(map_to)
+        self.register_load_state_dict_pre_hook(map_from, with_module=True)
+
+    def register_load_state_dict_pre_hook(self, hook, with_module=False):
+        if getattr(self, "load_state_dict_pre_hooks", None) is None:
+            self.load_state_dict_pre_hooks = OrderedDict()
+        handle = HookRemoveHelper(self.load_state_dict_pre_hooks)
+        self.load_state_dict_pre_hooks[handle._hook_id] = _WrappedHook(hook, self if with_module else None)
+        return handle
+
+    def set_state_dict(self, state_dict, use_structured_name: bool = True):
+        if getattr(self, "load_state_dict_pre_hooks", None) is not None:
+            for hook in self.load_state_dict_pre_hooks.values():
+                hook(state_dict)
+        return super().set_state_dict(self, state_dict, use_structured_name=use_structured_name)
+
+    load_dict = set_state_dict
+    set_dict = set_state_dict

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/models/__init__.py

@@ -0,0 +1,113 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+from ..utils import PPDIFFUSERS_SLOW_IMPORT, _LazyModule, is_paddle_available
+
+_import_structure = {}
+
+if is_paddle_available():
+    _import_structure["adapter"] = ["MultiAdapter", "T2IAdapter"]
+    _import_structure["autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
+    _import_structure["autoencoder_kl"] = ["AutoencoderKL"]
+    _import_structure["autoencoder_kl_cogvideox"] = ["AutoencoderKLCogVideoX"]
+    _import_structure["autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
+    _import_structure["autoencoder_tiny"] = ["AutoencoderTiny"]
+    _import_structure["consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
+    _import_structure["controlnet"] = ["ControlNetModel"]
+    _import_structure["dual_transformer_2d"] = ["DualTransformer2DModel"]
+    _import_structure["modeling_utils"] = ["ModelMixin"]
+    _import_structure["prior_transformer"] = ["PriorTransformer"]
+    _import_structure["t5_film_transformer"] = ["T5FilmDecoder"]
+    _import_structure["transformer_2d"] = ["Transformer2DModel"]
+    _import_structure["transformer_sd3"] = ["SD3Transformer2DModel"]
+    _import_structure["cogvideox_transformer_3d"] = ["CogVideoXTransformer3DModel"]
+    _import_structure["transformer_temporal"] = ["TransformerTemporalModel"]
+    _import_structure["unet_1d"] = ["UNet1DModel"]
+    _import_structure["unet_2d"] = ["UNet2DModel"]
+    _import_structure["unet_2d_condition"] = ["UNet2DConditionModel"]
+    _import_structure["unet_3d_condition"] = ["UNet3DConditionModel"]
+    _import_structure["unet_kandi3"] = ["Kandinsky3UNet"]
+    _import_structure["unet_motion_model"] = ["MotionAdapter", "UNetMotionModel"]
+    _import_structure["unet_spatio_temporal_condition"] = ["UNetSpatioTemporalConditionModel"]
+    _import_structure["vq_model"] = ["VQModel"]
+    _import_structure["uvit_t2i"] = ["UViTT2IModel"]
+    _import_structure["dit_llama"] = ["DiTLLaMA2DModel"]
+    _import_structure["dit_llama_t2i"] = ["DiTLLaMAT2IModel"]
+    # NOTE, new add
+    _import_structure["lvdm_vae"] = ["LVDMAutoencoderKL"]
+    _import_structure["lvdm_unet_3d"] = ["LVDMUNet3DModel"]
+    _import_structure["ema"] = ["LitEma"]
+    _import_structure["paddleinfer_runtime"] = ["PaddleInferRuntimeModel"]
+    # NOTE, new add
+    _import_structure["modelscope_autoencoder_img2vid"] = ["AutoencoderKL_imgtovideo"]
+    _import_structure["modelscope_gaussian_diffusion"] = ["GaussianDiffusion"]
+    _import_structure["modelscope_gaussion_sdedit"] = ["GaussianDiffusion_SDEdit"]
+    _import_structure["modelscope_st_unet"] = ["STUNetModel"]
+    _import_structure["modelscope_st_unet_video2video"] = ["Vid2VidSTUNet"]
+    # NOTE, new add
+    _import_structure["controlnet_sd3"] = ["SD3ControlNetModel", "SD3MultiControlNetModel"]
+    # NOTE, new add
+    _import_structure["vctrl"] = ["VCtrlModel"]
+    _import_structure["cogvideox_transformer_3d_vctrl"] = ["CogVideoXTransformer3DVCtrlModel"]
+
+
+if TYPE_CHECKING or PPDIFFUSERS_SLOW_IMPORT:
+    if is_paddle_available():
+        from .adapter import MultiAdapter, T2IAdapter
+        from .autoencoder_asym_kl import AsymmetricAutoencoderKL
+        from .autoencoder_kl import AutoencoderKL
+        from .autoencoder_kl_cogvideox import AutoencoderKLCogVideoX
+        from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
+        from .autoencoder_tiny import AutoencoderTiny
+        from .cogvideox_transformer_3d import CogVideoXTransformer3DModel
+        from .cogvideox_transformer_3d_vctrl import CogVideoXTransformer3DVCtrlModel
+        from .consistency_decoder_vae import ConsistencyDecoderVAE
+        from .controlnet import ControlNetModel
+        from .controlnet_sd3 import SD3ControlNetModel, SD3MultiControlNetModel
+        from .dit_llama import DiTLLaMA2DModel
+        from .dit_llama_t2i import DiTLLaMAT2IModel
+        from .dual_transformer_2d import DualTransformer2DModel
+
+        # NOTE, new add
+        from .ema import LitEma
+        from .lvdm_unet_3d import LVDMUNet3DModel
+        from .lvdm_vae import LVDMAutoencoderKL
+        from .modeling_utils import ModelMixin
+        from .modelscope_autoencoder_img2vid import AutoencoderKL_imgtovideo
+        from .modelscope_gaussian_diffusion import GaussianDiffusion
+        from .modelscope_gaussion_sdedit import GaussianDiffusion_SDEdit
+        from .modelscope_st_unet import STUNetModel
+        from .modelscope_st_unet_video2video import Vid2VidSTUNet
+        from .paddleinfer_runtime import PaddleInferRuntimeModel
+        from .prior_transformer import PriorTransformer
+        from .t5_film_transformer import T5FilmDecoder
+        from .transformer_2d import Transformer2DModel
+        from .transformer_sd3 import SD3Transformer2DModel
+        from .transformer_temporal import TransformerTemporalModel
+        from .unet_1d import UNet1DModel
+        from .unet_2d import UNet2DModel
+        from .unet_2d_condition import UNet2DConditionModel
+        from .unet_3d_condition import UNet3DConditionModel
+        from .unet_kandi3 import Kandinsky3UNet
+        from .unet_motion_model import MotionAdapter, UNetMotionModel
+        from .unet_spatio_temporal_condition import UNetSpatioTemporalConditionModel
+        from .uvit_t2i import UViTT2IModel
+        from .vctrl import VCtrlModel
+        from .vq_model import VQModel
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/models/activations.py

@@ -0,0 +1,123 @@
+# coding=utf-8
+# Copyright 2023 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle
+import paddle.nn.functional as F
+from paddle import nn
+
+from ..utils import USE_PEFT_BACKEND
+from .lora import LoRACompatibleLinear
+
+ACTIVATION_FUNCTIONS = {
+    "swish": nn.Silu(),
+    "silu": nn.Silu(),
+    "mish": nn.Mish(),
+    "gelu": nn.GELU(),
+    "relu": nn.ReLU(),
+}
+
+
+def get_activation(act_fn: str) -> nn.Layer:
+    """Helper function to get activation function from string.
+
+    Args:
+        act_fn (str): Name of activation function.
+
+    Returns:
+        nn.Layer: Activation function.
+    """
+
+    act_fn = act_fn.lower()
+    if act_fn in ACTIVATION_FUNCTIONS:
+        return ACTIVATION_FUNCTIONS[act_fn]
+    else:
+        raise ValueError(f"Unsupported activation function: {act_fn}")
+
+class FP32SiLU(nn.Layer):
+    r"""
+    SiLU activation function with input upcasted to paddle.float32.
+    """
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, inputs: paddle.Tensor) -> paddle.Tensor:
+        return F.silu(inputs.astype(paddle.float32)).astype(inputs.dtype)
+
+class GELU(nn.Layer):
+    r"""
+    GELU activation function with tanh approximation support with `approximate="tanh"`.
+
+    Parameters:
+        dim_in (`int`): The number of channels in the input.
+        dim_out (`int`): The number of channels in the output.
+        approximate (`str`, *optional*, defaults to `"none"`): If `"tanh"`, use tanh approximation.
+    """
+
+    def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool=True):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out, bias_attr=bias)
+        self.approximate = approximate
+
+    def gelu(self, gate: paddle.Tensor) -> paddle.Tensor:
+        return F.gelu(gate, approximate=self.approximate != "none")
+
+    def forward(self, hidden_states):
+        hidden_states = self.proj(hidden_states)
+        hidden_states = self.gelu(hidden_states)
+        return hidden_states
+
+
+class GEGLU(nn.Layer):
+    r"""
+    A [variant](https://arxiv.org/abs/2002.05202) of the gated linear unit activation function.
+
+    Parameters:
+        dim_in (`int`): The number of channels in the input.
+        dim_out (`int`): The number of channels in the output.
+    """
+
+    def __init__(self, dim_in: int, dim_out: int, bias: bool=True):
+        super().__init__()
+        linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
+
+        self.proj = linear_cls(dim_in, dim_out * 2, bias_attr=bias)
+
+    def gelu(self, gate: paddle.Tensor) -> paddle.Tensor:
+        return F.gelu(gate)
+
+    def forward(self, hidden_states, scale: float = 1.0):
+        args = () if USE_PEFT_BACKEND else (scale,)
+        hidden_states, gate = self.proj(hidden_states, *args).chunk(2, axis=-1)
+        return hidden_states * self.gelu(gate)
+
+
+class ApproximateGELU(nn.Layer):
+    r"""
+    The approximate form of the Gaussian Error Linear Unit (GELU). For more details, see section 2 of this
+    [paper](https://arxiv.org/abs/1606.08415).
+
+    Parameters:
+        dim_in (`int`): The number of channels in the input.
+        dim_out (`int`): The number of channels in the output.
+    """
+
+    def __init__(self, dim_in: int, dim_out: int, bias: bool=True):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out, bias_attr=bias)
+
+    def forward(self, x: paddle.Tensor) -> paddle.Tensor:
+        x = self.proj(x)
+        return x * F.sigmoid(1.702 * x)

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/models/adapter.py

@@ -0,0 +1,583 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+from typing import Callable, List, Optional, Union
+
+import paddle
+import paddle.nn as nn
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import logging
+from .modeling_utils import ModelMixin
+
+logger = logging.get_logger(__name__)
+
+
+class MultiAdapter(ModelMixin):
+    r"""
+    MultiAdapter is a wrapper model that contains multiple adapter models and merges their outputs according to
+    user-assigned weighting.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library
+    implements for all the model (such as downloading or saving, etc.)
+
+    Parameters:
+        adapters (`List[T2IAdapter]`, *optional*, defaults to None):
+            A list of `T2IAdapter` model instances.
+    """
+
+    def __init__(self, adapters: List["T2IAdapter"]):
+        super(MultiAdapter, self).__init__()
+
+        self.num_adapter = len(adapters)
+        self.adapters = nn.LayerList(adapters)
+
+        if len(adapters) == 0:
+            raise ValueError("Expecting at least one adapter")
+
+        if len(adapters) == 1:
+            raise ValueError("For a single adapter, please use the `T2IAdapter` class instead of `MultiAdapter`")
+
+        # The outputs from each adapter are added together with a weight.
+        # This means that the change in dimensions from downsampling must
+        # be the same for all adapters. Inductively, it also means the
+        # downscale_factor and total_downscale_factor must be the same for all
+        # adapters.
+        first_adapter_total_downscale_factor = adapters[0].total_downscale_factor
+        first_adapter_downscale_factor = adapters[0].downscale_factor
+        for idx in range(1, len(adapters)):
+            if (
+                adapters[idx].total_downscale_factor != first_adapter_total_downscale_factor
+                or adapters[idx].downscale_factor != first_adapter_downscale_factor
+            ):
+                raise ValueError(
+                    f"Expecting all adapters to have the same downscaling behavior, but got:\n"
+                    f"adapters[0].total_downscale_factor={first_adapter_total_downscale_factor}\n"
+                    f"adapters[0].downscale_factor={first_adapter_downscale_factor}\n"
+                    f"adapter[`{idx}`].total_downscale_factor={adapters[idx].total_downscale_factor}\n"
+                    f"adapter[`{idx}`].downscale_factor={adapters[idx].downscale_factor}"
+                )
+
+        self.total_downscale_factor = first_adapter_total_downscale_factor
+        self.downscale_factor = first_adapter_downscale_factor
+
+    def forward(self, xs: paddle.Tensor, adapter_weights: Optional[List[float]] = None) -> List[paddle.Tensor]:
+        r"""
+        Args:
+            xs (`paddle.Tensor`):
+                (batch, channel, height, width) input images for multiple adapter models concated along dimension 1,
+                `channel` should equal to `num_adapter` * "number of channel of image".
+            adapter_weights (`List[float]`, *optional*, defaults to None):
+                List of floats representing the weight which will be multiply to each adapter's output before adding
+                them together.
+        """
+        if adapter_weights is None:
+            adapter_weights = paddle.to_tensor([1 / self.num_adapter] * self.num_adapter)
+        else:
+            adapter_weights = paddle.to_tensor(adapter_weights)
+
+        accume_state = None
+        for x, w, adapter in zip(xs, adapter_weights, self.adapters):
+            features = adapter(x)
+            if accume_state is None:
+                accume_state = features
+                for i in range(len(accume_state)):
+                    accume_state[i] = w * accume_state[i]
+            else:
+                for i in range(len(features)):
+                    accume_state[i] += w * features[i]
+        return accume_state
+
+    def save_pretrained(
+        self,
+        save_directory: Union[str, os.PathLike],
+        is_main_process: bool = True,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+        variant: Optional[str] = None,
+    ):
+        """
+        Save a model and its configuration file to a directory, so that it can be re-loaded using the
+        `[`~models.adapter.MultiAdapter.from_pretrained`]` class method.
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to which to save. Will be created if it doesn't exist.
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful when in distributed training like
+                TPUs and need to call this function on all processes. In this case, set `is_main_process=True` only on
+                the main process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful on distributed training like TPUs when one
+                need to replace `torch.save` by another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+            variant (`str`, *optional*):
+                If specified, weights are saved in the format pytorch_model.<variant>.bin.
+        """
+        idx = 0
+        model_path_to_save = save_directory
+        for adapter in self.adapters:
+            adapter.save_pretrained(
+                model_path_to_save,
+                is_main_process=is_main_process,
+                save_function=save_function,
+                safe_serialization=safe_serialization,
+                variant=variant,
+            )
+
+            idx += 1
+            model_path_to_save = model_path_to_save + f"_{idx}"
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_path: Optional[Union[str, os.PathLike]], **kwargs):
+        r"""
+        Instantiate a pretrained MultiAdapter model from multiple pre-trained adapter models.
+
+        The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated). To train
+        the model, you should first set it back in training mode with `model.train()`.
+
+        The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
+        pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
+        task.
+
+        The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
+        weights are discarded.
+
+        Parameters:
+            pretrained_model_path (`os.PathLike`):
+                A path to a *directory* containing model weights saved using
+                [`~ppdiffusers.models.adapter.MultiAdapter.save_pretrained`], e.g., `./my_model_directory/adapter`.
+            torch_dtype (`str` or `torch.dtype`, *optional*):
+                Override the default `torch.dtype` and load the model under this dtype. If `"auto"` is passed the dtype
+                will be automatically derived from the model's weights.
+            output_loading_info(`bool`, *optional*, defaults to `False`):
+                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+            device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
+                A map that specifies where each submodule should go. It doesn't need to be refined to each
+                parameter/buffer name, once a given module name is inside, every submodule of it will be sent to the
+                same device.
+
+                To have Accelerate compute the most optimized `device_map` automatically, set `device_map="auto"`. For
+                more information about each option see [designing a device
+                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
+            max_memory (`Dict`, *optional*):
+                A dictionary device identifier to maximum memory. Will default to the maximum memory available for each
+                GPU and the available CPU RAM if unset.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading by not initializing the weights and only loading the pre-trained weights. This
+                also tries to not use more than 1x model size in CPU memory (including peak memory) while loading the
+                model. This is only supported when torch version >= 1.9.0. If you are using an older version of torch,
+                setting this argument to `True` will raise an error.
+            variant (`str`, *optional*):
+                If specified load weights from `variant` filename, *e.g.* pytorch_model.<variant>.bin. `variant` is
+                ignored when using `from_flax`.
+            use_safetensors (`bool`, *optional*, defaults to `None`):
+                If set to `None`, the `safetensors` weights will be downloaded if they're available **and** if the
+                `safetensors` library is installed. If set to `True`, the model will be forcibly loaded from
+                `safetensors` weights. If set to `False`, loading will *not* use `safetensors`.
+        """
+        idx = 0
+        adapters = []
+
+        # load adapter and append to list until no adapter directory exists anymore
+        # first adapter has to be saved under `./mydirectory/adapter` to be compliant with `DiffusionPipeline.from_pretrained`
+        # second, third, ... adapters have to be saved under `./mydirectory/adapter_1`, `./mydirectory/adapter_2`, ...
+        model_path_to_load = pretrained_model_path
+        while os.path.isdir(model_path_to_load):
+            adapter = T2IAdapter.from_pretrained(model_path_to_load, **kwargs)
+            adapters.append(adapter)
+
+            idx += 1
+            model_path_to_load = pretrained_model_path + f"_{idx}"
+
+        logger.info(f"{len(adapters)} adapters loaded from {pretrained_model_path}.")
+
+        if len(adapters) == 0:
+            raise ValueError(
+                f"No T2IAdapters found under {os.path.dirname(pretrained_model_path)}. Expected at least {pretrained_model_path + '_0'}."
+            )
+
+        return cls(adapters)
+
+
+class T2IAdapter(ModelMixin, ConfigMixin):
+    r"""
+    A simple ResNet-like model that accepts images containing control signals such as keyposes and depth. The model
+    generates multiple feature maps that are used as additional conditioning in [`UNet2DConditionModel`]. The model's
+    architecture follows the original implementation of
+    [Adapter](https://github.com/TencentARC/T2I-Adapter/blob/686de4681515662c0ac2ffa07bf5dda83af1038a/ldm/modules/encoders/adapter.py#L97)
+     and
+     [AdapterLight](https://github.com/TencentARC/T2I-Adapter/blob/686de4681515662c0ac2ffa07bf5dda83af1038a/ldm/modules/encoders/adapter.py#L235).
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library
+    implements for all the model (such as downloading or saving, etc.)
+
+    Parameters:
+        in_channels (`int`, *optional*, defaults to 3):
+            Number of channels of Adapter's input(*control image*). Set this parameter to 1 if you're using gray scale
+            image as *control image*.
+        channels (`List[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The number of channel of each downsample block's output hidden state. The `len(block_out_channels)` will
+            also determine the number of downsample blocks in the Adapter.
+        num_res_blocks (`int`, *optional*, defaults to 2):
+            Number of ResNet blocks in each downsample block.
+        downscale_factor (`int`, *optional*, defaults to 8):
+            A factor that determines the total downscale factor of the Adapter.
+        adapter_type (`str`, *optional*, defaults to `full_adapter`):
+            The type of Adapter to use. Choose either `full_adapter` or `full_adapter_xl` or `light_adapter`.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 3,
+        channels: List[int] = [320, 640, 1280, 1280],
+        num_res_blocks: int = 2,
+        downscale_factor: int = 8,
+        adapter_type: str = "full_adapter",
+    ):
+        super().__init__()
+
+        if adapter_type == "full_adapter":
+            self.adapter = FullAdapter(in_channels, channels, num_res_blocks, downscale_factor)
+        elif adapter_type == "full_adapter_xl":
+            self.adapter = FullAdapterXL(in_channels, channels, num_res_blocks, downscale_factor)
+        elif adapter_type == "light_adapter":
+            self.adapter = LightAdapter(in_channels, channels, num_res_blocks, downscale_factor)
+        else:
+            raise ValueError(
+                f"Unsupported adapter_type: '{adapter_type}'. Choose either 'full_adapter' or "
+                "'full_adapter_xl' or 'light_adapter'."
+            )
+
+    def forward(self, x: paddle.Tensor) -> List[paddle.Tensor]:
+        r"""
+        This function processes the input tensor `x` through the adapter model and returns a list of feature tensors,
+        each representing information extracted at a different scale from the input. The length of the list is
+        determined by the number of downsample blocks in the Adapter, as specified by the `channels` and
+        `num_res_blocks` parameters during initialization.
+        """
+        return self.adapter(x)
+
+    @property
+    def total_downscale_factor(self):
+        return self.adapter.total_downscale_factor
+
+    @property
+    def downscale_factor(self):
+        """The downscale factor applied in the T2I-Adapter's initial pixel unshuffle operation. If an input image's dimensions are
+        not evenly divisible by the downscale_factor then an exception will be raised.
+        """
+        return self.adapter.unshuffle.downscale_factor
+
+
+# full adapter
+
+
+class FullAdapter(nn.Layer):
+    r"""
+    See [`T2IAdapter`] for more information.
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        channels: List[int] = [320, 640, 1280, 1280],
+        num_res_blocks: int = 2,
+        downscale_factor: int = 8,
+    ):
+        super().__init__()
+
+        in_channels = in_channels * downscale_factor**2
+
+        self.unshuffle = nn.PixelUnshuffle(downscale_factor)
+        self.conv_in = nn.Conv2D(in_channels, channels[0], kernel_size=3, padding=1)
+
+        self.body = nn.LayerList(
+            [
+                AdapterBlock(channels[0], channels[0], num_res_blocks),
+                *[
+                    AdapterBlock(channels[i - 1], channels[i], num_res_blocks, down=True)
+                    for i in range(1, len(channels))
+                ],
+            ]
+        )
+
+        self.total_downscale_factor = downscale_factor * 2 ** (len(channels) - 1)
+
+    def forward(self, x: paddle.Tensor) -> List[paddle.Tensor]:
+        r"""
+        This method processes the input tensor `x` through the FullAdapter model and performs operations including
+        pixel unshuffling, convolution, and a stack of AdapterBlocks. It returns a list of feature tensors, each
+        capturing information at a different stage of processing within the FullAdapter model. The number of feature
+        tensors in the list is determined by the number of downsample blocks specified during initialization.
+        """
+        x = self.unshuffle(x)
+        x = self.conv_in(x)
+
+        features = []
+
+        for block in self.body:
+            x = block(x)
+            features.append(x)
+
+        return features
+
+
+class FullAdapterXL(nn.Layer):
+    r"""
+    See [`T2IAdapter`] for more information.
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        channels: List[int] = [320, 640, 1280, 1280],
+        num_res_blocks: int = 2,
+        downscale_factor: int = 16,
+    ):
+        super().__init__()
+
+        in_channels = in_channels * downscale_factor**2
+
+        self.unshuffle = nn.PixelUnshuffle(downscale_factor)
+        self.conv_in = nn.Conv2D(in_channels, channels[0], kernel_size=3, padding=1)
+
+        self.body = []
+        # blocks to extract XL features with dimensions of [320, 64, 64], [640, 64, 64], [1280, 32, 32], [1280, 32, 32]
+        for i in range(len(channels)):
+            if i == 1:
+                self.body.append(AdapterBlock(channels[i - 1], channels[i], num_res_blocks))
+            elif i == 2:
+                self.body.append(AdapterBlock(channels[i - 1], channels[i], num_res_blocks, down=True))
+            else:
+                self.body.append(AdapterBlock(channels[i], channels[i], num_res_blocks))
+
+        self.body = nn.LayerList(self.body)
+        # XL has only one downsampling AdapterBlock.
+        self.total_downscale_factor = downscale_factor * 2
+
+    def forward(self, x: paddle.Tensor) -> List[paddle.Tensor]:
+        r"""
+        This method takes the tensor x as input and processes it through FullAdapterXL model. It consists of operations
+        including unshuffling pixels, applying convolution layer and appending each block into list of feature tensors.
+        """
+        x = self.unshuffle(x)
+        x = self.conv_in(x)
+
+        features = []
+
+        for block in self.body:
+            x = block(x)
+            features.append(x)
+
+        return features
+
+
+class AdapterBlock(nn.Layer):
+    r"""
+    An AdapterBlock is a helper model that contains multiple ResNet-like blocks. It is used in the `FullAdapter` and
+    `FullAdapterXL` models.
+
+    Parameters:
+        in_channels (`int`):
+            Number of channels of AdapterBlock's input.
+        out_channels (`int`):
+            Number of channels of AdapterBlock's output.
+        num_res_blocks (`int`):
+            Number of ResNet blocks in the AdapterBlock.
+        down (`bool`, *optional*, defaults to `False`):
+            Whether to perform downsampling on AdapterBlock's input.
+    """
+
+    def __init__(self, in_channels: int, out_channels: int, num_res_blocks: int, down: bool = False):
+        super().__init__()
+
+        self.downsample = None
+        if down:
+            self.downsample = nn.AvgPool2D(kernel_size=2, stride=2, ceil_mode=True)
+
+        self.in_conv = None
+        if in_channels != out_channels:
+            self.in_conv = nn.Conv2D(in_channels, out_channels, kernel_size=1)
+
+        self.resnets = nn.Sequential(
+            *[AdapterResnetBlock(out_channels) for _ in range(num_res_blocks)],
+        )
+
+    def forward(self, x: paddle.Tensor) -> paddle.Tensor:
+        r"""
+        This method takes tensor x as input and performs operations downsampling and convolutional layers if the
+        self.downsample and self.in_conv properties of AdapterBlock model are specified. Then it applies a series of
+        residual blocks to the input tensor.
+        """
+        if self.downsample is not None:
+            x = self.downsample(x)
+
+        if self.in_conv is not None:
+            x = self.in_conv(x)
+
+        x = self.resnets(x)
+
+        return x
+
+
+class AdapterResnetBlock(nn.Layer):
+    r"""
+    An `AdapterResnetBlock` is a helper model that implements a ResNet-like block.
+
+    Parameters:
+        channels (`int`):
+            Number of channels of AdapterResnetBlock's input and output.
+    """
+
+    def __init__(self, channels: int):
+        super().__init__()
+        self.block1 = nn.Conv2D(channels, channels, kernel_size=3, padding=1)
+        self.act = nn.ReLU()
+        self.block2 = nn.Conv2D(channels, channels, kernel_size=1)
+
+    def forward(self, x: paddle.Tensor) -> paddle.Tensor:
+        r"""
+        This method takes input tensor x and applies a convolutional layer, ReLU activation, and another convolutional
+        layer on the input tensor. It returns addition with the input tensor.
+        """
+
+        h = self.act(self.block1(x))
+        h = self.block2(h)
+
+        return h + x
+
+
+# light adapter
+
+
+class LightAdapter(nn.Layer):
+    r"""
+    See [`T2IAdapter`] for more information.
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        channels: List[int] = [320, 640, 1280],
+        num_res_blocks: int = 4,
+        downscale_factor: int = 8,
+    ):
+        super().__init__()
+
+        in_channels = in_channels * downscale_factor**2
+
+        self.unshuffle = nn.PixelUnshuffle(downscale_factor)
+
+        self.body = nn.LayerList(
+            [
+                LightAdapterBlock(in_channels, channels[0], num_res_blocks),
+                *[
+                    LightAdapterBlock(channels[i], channels[i + 1], num_res_blocks, down=True)
+                    for i in range(len(channels) - 1)
+                ],
+                LightAdapterBlock(channels[-1], channels[-1], num_res_blocks, down=True),
+            ]
+        )
+
+        self.total_downscale_factor = downscale_factor * (2 ** len(channels))
+
+    def forward(self, x: paddle.Tensor) -> List[paddle.Tensor]:
+        r"""
+        This method takes the input tensor x and performs downscaling and appends it in list of feature tensors. Each
+        feature tensor corresponds to a different level of processing within the LightAdapter.
+        """
+        x = self.unshuffle(x)
+
+        features = []
+
+        for block in self.body:
+            x = block(x)
+            features.append(x)
+
+        return features
+
+
+class LightAdapterBlock(nn.Layer):
+    r"""
+    A `LightAdapterBlock` is a helper model that contains multiple `LightAdapterResnetBlocks`. It is used in the
+    `LightAdapter` model.
+
+    Parameters:
+        in_channels (`int`):
+            Number of channels of LightAdapterBlock's input.
+        out_channels (`int`):
+            Number of channels of LightAdapterBlock's output.
+        num_res_blocks (`int`):
+            Number of LightAdapterResnetBlocks in the LightAdapterBlock.
+        down (`bool`, *optional*, defaults to `False`):
+            Whether to perform downsampling on LightAdapterBlock's input.
+    """
+
+    def __init__(self, in_channels: int, out_channels: int, num_res_blocks: int, down: bool = False):
+        super().__init__()
+        mid_channels = out_channels // 4
+
+        self.downsample = None
+        if down:
+            self.downsample = nn.AvgPool2D(kernel_size=2, stride=2, ceil_mode=True)
+
+        self.in_conv = nn.Conv2D(in_channels, mid_channels, kernel_size=1)
+        self.resnets = nn.Sequential(*[LightAdapterResnetBlock(mid_channels) for _ in range(num_res_blocks)])
+        self.out_conv = nn.Conv2D(mid_channels, out_channels, kernel_size=1)
+
+    def forward(self, x: paddle.Tensor) -> paddle.Tensor:
+        r"""
+        This method takes tensor x as input and performs downsampling if required. Then it applies in convolution
+        layer, a sequence of residual blocks, and out convolutional layer.
+        """
+        if self.downsample is not None:
+            x = self.downsample(x)
+
+        x = self.in_conv(x)
+        x = self.resnets(x)
+        x = self.out_conv(x)
+
+        return x
+
+
+class LightAdapterResnetBlock(nn.Layer):
+    """
+    A `LightAdapterResnetBlock` is a helper model that implements a ResNet-like block with a slightly different
+    architecture than `AdapterResnetBlock`.
+
+    Parameters:
+        channels (`int`):
+            Number of channels of LightAdapterResnetBlock's input and output.
+    """
+
+    def __init__(self, channels: int):
+        super().__init__()
+        self.block1 = nn.Conv2D(channels, channels, kernel_size=3, padding=1)
+        self.act = nn.ReLU()
+        self.block2 = nn.Conv2D(channels, channels, kernel_size=3, padding=1)
+
+    def forward(self, x: paddle.Tensor) -> paddle.Tensor:
+        r"""
+        This function takes input tensor x and processes it through one convolutional layer, ReLU activation, and
+        another convolutional layer and adds it to input tensor.
+        """
+
+        h = self.act(self.block1(x))
+        h = self.block2(h)
+
+        return h + x

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/models/attention.py

@@ -0,0 +1,721 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any, Dict, Optional
+
+import paddle
+from paddle import nn
+import paddle.nn.functional as F
+
+from ..utils import USE_PEFT_BACKEND
+from ..utils.paddle_utils import maybe_allow_in_graph
+from .activations import GEGLU, GELU, ApproximateGELU
+from .attention_processor import Attention, JointAttnProcessor2_5
+from .embeddings import SinusoidalPositionalEmbedding
+from .lora import LoRACompatibleLinear
+from .normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, SD35AdaLayerNormZeroX
+
+
+def _chunked_feed_forward(
+    ff: nn.Layer, hidden_states: paddle.Tensor, chunk_dim: int, chunk_size: int, lora_scale: Optional[float] = None
+):
+    # "feed_forward_chunk_size" can be used to save memory
+    if hidden_states.shape[chunk_dim] % chunk_size != 0:
+        raise ValueError(
+            f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+        )
+
+    num_chunks = hidden_states.shape[chunk_dim] // chunk_size
+    if lora_scale is None:
+        ff_output = paddle.concat(
+            [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, axis=chunk_dim)],
+            dim=chunk_dim,
+        )
+    else:
+        # TOOD(Patrick): LoRA scale can be removed once PEFT refactor is complete
+        ff_output = paddle.concat(
+            [ff(hid_slice, scale=lora_scale) for hid_slice in hidden_states.chunk(num_chunks, axis=chunk_dim)],
+            axis=chunk_dim,
+        )
+
+    return ff_output
+
+
+@maybe_allow_in_graph
+class GatedSelfAttentionDense(nn.Layer):
+    r"""
+    A gated self-attention dense layer that combines visual features and object features.
+
+    Parameters:
+        query_dim (`int`): The number of channels in the query.
+        context_dim (`int`): The number of channels in the context.
+        n_heads (`int`): The number of heads to use for attention.
+        d_head (`int`): The number of channels in each head.
+    """
+
+    def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
+        super().__init__()
+
+        # we need a linear projection since we need cat visual feature and obj feature
+        self.linear = nn.Linear(context_dim, query_dim)
+
+        self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
+        self.ff = FeedForward(query_dim, activation_fn="geglu")
+
+        self.norm1 = nn.LayerNorm(query_dim)
+        self.norm2 = nn.LayerNorm(query_dim)
+
+        self.alpha_attn = nn.Parameter(paddle.to_tensor(0.0))
+        self.alpha_dense = nn.Parameter(paddle.to_tensor(0.0))
+
+        self.enabled = True
+
+    def forward(self, x: paddle.Tensor, objs: paddle.Tensor) -> paddle.Tensor:
+        if not self.enabled:
+            return x
+
+        n_visual = x.shape[1]
+        objs = self.linear(objs)
+
+        x = x + self.alpha_attn.tanh() * self.attn(self.norm1(paddle.concat([x, objs], axis=1)))[:, :n_visual, :]
+        x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
+
+        return x
+
+@maybe_allow_in_graph
+class JointTransformerBlock(nn.Layer):
+    r"""
+    A Transformer block following the MMDiT architecture, introduced in Stable Diffusion 3.
+    Reference: https://arxiv.org/abs/2403.03206
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        context_pre_only (`bool`): Boolean to determine if we should add some blocks associated with the
+            processing of `context` conditions.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        context_pre_only: bool = False,
+        qk_norm: Optional[str] = None,
+        use_dual_attention: bool = False,
+    ):
+        super().__init__()
+
+        self.use_dual_attention = use_dual_attention
+        self.context_pre_only = context_pre_only
+        context_norm_type = "ada_norm_continous" if context_pre_only else "ada_norm_zero"
+
+        if use_dual_attention:
+            self.norm1 = SD35AdaLayerNormZeroX(dim)
+        else:
+            self.norm1 = AdaLayerNormZero(dim)
+        
+        if context_norm_type == "ada_norm_continous":
+            self.norm1_context = AdaLayerNormContinuous(
+                dim, dim, elementwise_affine=False, eps=1e-6, bias=True, norm_type="layer_norm"
+            )
+        elif context_norm_type == "ada_norm_zero":
+            self.norm1_context = AdaLayerNormZero(dim)
+        else:
+            raise ValueError(
+                f"Unknown context_norm_type: {context_norm_type}, currently only support `ada_norm_continous`, `ada_norm_zero`"
+            )
+        
+        if hasattr(F, "scaled_dot_product_attention"):
+            processor = JointAttnProcessor2_5()
+        else:
+            raise ValueError(
+                "The current PyTorch version does not support the `scaled_dot_product_attention` function."
+            )
+        
+        self.attn = Attention(
+            query_dim=dim,
+            cross_attention_dim=None,
+            added_kv_proj_dim=dim,
+            dim_head=attention_head_dim // num_attention_heads,
+            heads=num_attention_heads,
+            out_dim=attention_head_dim,
+            context_pre_only=context_pre_only,
+            bias=True,
+            processor=processor,
+            qk_norm=qk_norm,
+            eps=1e-6,
+        )
+
+        if use_dual_attention:
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=None,
+                dim_head=attention_head_dim // num_attention_heads,
+                heads=num_attention_heads,
+                out_dim=attention_head_dim,
+                bias=True,
+                processor=processor,
+                qk_norm=qk_norm,
+                eps=1e-6,
+            )
+        else:
+            self.attn2 = None
+
+        self.norm2 = nn.LayerNorm(dim, weight_attr=False, bias_attr=False, epsilon=1e-6)
+        self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
+
+        if not context_pre_only:
+            self.norm2_context = nn.LayerNorm(dim, weight_attr=False, bias_attr=False, epsilon=1e-6)
+            self.ff_context = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
+        else:
+            self.norm2_context = None
+            self.ff_context = None
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    # Copied from diffusers.models.attention.BasicTransformerBlock.set_chunk_feed_forward
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self, hidden_states: paddle.Tensor, encoder_hidden_states: paddle.Tensor, temb: paddle.Tensor
+    ):
+        if self.use_dual_attention:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp, norm_hidden_states2, gate_msa2 = self.norm1(
+                hidden_states, emb=temb
+            )
+        else:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
+
+        if self.context_pre_only:
+            norm_encoder_hidden_states = self.norm1_context(encoder_hidden_states, temb)
+        else:
+            norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
+                encoder_hidden_states, emb=temb
+            )
+
+        # Attention.
+        attn_output, context_attn_output = self.attn(
+            hidden_states=norm_hidden_states, encoder_hidden_states=norm_encoder_hidden_states, 
+        )
+
+        # Process attention outputs for the `hidden_states`.
+        attn_output = gate_msa.unsqueeze(1) * attn_output
+        hidden_states = hidden_states + attn_output
+
+        if self.use_dual_attention:
+            attn_output2 = self.attn2(hidden_states=norm_hidden_states2)
+            attn_output2 = gate_msa2.unsqueeze(1) * attn_output2
+            hidden_states = hidden_states + attn_output2
+
+        norm_hidden_states = self.norm2(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+        ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+        hidden_states = hidden_states + ff_output
+
+        # Process attention outputs for the `encoder_hidden_states`.
+        if self.context_pre_only:
+            encoder_hidden_states = None
+        else:
+            context_attn_output = c_gate_msa.unsqueeze(1) * context_attn_output
+            encoder_hidden_states = encoder_hidden_states + context_attn_output
+
+            norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
+            norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
+            if self._chunk_size is not None:
+                # "feed_forward_chunk_size" can be used to save memory
+                context_ff_output = _chunked_feed_forward(
+                    self.ff_context, norm_encoder_hidden_states, self._chunk_dim, self._chunk_size
+                )
+            else:
+                context_ff_output = self.ff_context(norm_encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
+
+        return encoder_hidden_states, hidden_states
+
+@maybe_allow_in_graph
+class BasicTransformerBlock(nn.Layer):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, *optional*, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*, defaults to `None`):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+
+        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positional_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_zero:
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        else:
+            norm_elementwise_affine_kwargs = (
+                {} if norm_elementwise_affine else dict(weight_attr=False, bias_attr=False)
+            )
+            self.norm1 = nn.LayerNorm(dim, epsilon=norm_eps, **norm_elementwise_affine_kwargs)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            norm_elementwise_affine_kwargs = (
+                {} if norm_elementwise_affine else dict(weight_attr=False, bias_attr=False)
+            )
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim, epsilon=norm_eps, **norm_elementwise_affine_kwargs)
+            )
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            norm_elementwise_affine_kwargs = (
+                {} if norm_elementwise_affine else dict(weight_attr=False, bias_attr=False)
+            )
+            self.norm3 = nn.LayerNorm(dim, epsilon=norm_eps, **norm_elementwise_affine_kwargs)
+
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+        )
+
+        # 4. Fuser
+        if attention_type == "gated" or attention_type == "gated-text-image":
+            self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
+
+        # 5. Scale-shift for PixArt-Alpha.
+        if self.use_ada_layer_norm_single:
+            self.scale_shift_table = nn.Parameter(paddle.randn([6, dim]) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: paddle.Tensor,
+        attention_mask: Optional[paddle.Tensor] = None,
+        encoder_hidden_states: Optional[paddle.Tensor] = None,
+        encoder_attention_mask: Optional[paddle.Tensor] = None,
+        timestep: Optional[paddle.Tensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[paddle.Tensor] = None,
+    ) -> paddle.Tensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        if self.use_ada_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.use_ada_layer_norm_zero:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        elif self.use_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states)
+        elif self.use_ada_layer_norm_single:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape([batch_size, 6, -1])
+            ).chunk(6, axis=1)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+            norm_hidden_states = norm_hidden_states.squeeze(1)
+        else:
+            raise ValueError("Incorrect norm used")
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. Retrieve lora scale.
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+        # 2. Prepare GLIGEN inputs
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.use_ada_layer_norm_single:
+            attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 2.5 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            if self.use_ada_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero or self.use_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states)
+            elif self.use_ada_layer_norm_single:
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            else:
+                raise ValueError("Incorrect norm")
+
+            if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm3(hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+        if self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm2(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            ff_output = _chunked_feed_forward(
+                self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size, lora_scale=lora_scale
+            )
+        else:
+            ff_output = self.ff(norm_hidden_states, scale=lora_scale)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.use_ada_layer_norm_single:
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+@maybe_allow_in_graph
+class TemporalBasicTransformerBlock(nn.Layer):
+    r"""
+    A basic Transformer block for video like data.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        time_mix_inner_dim (`int`): The number of channels for temporal attention.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        time_mix_inner_dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        cross_attention_dim: Optional[int] = None,
+    ):
+        super().__init__()
+        self.is_res = dim == time_mix_inner_dim
+
+        self.norm_in = nn.LayerNorm(dim)
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        self.norm_in = nn.LayerNorm(dim)
+        self.ff_in = FeedForward(
+            dim,
+            dim_out=time_mix_inner_dim,
+            activation_fn="geglu",
+        )
+
+        self.norm1 = nn.LayerNorm(time_mix_inner_dim)
+        self.attn1 = Attention(
+            query_dim=time_mix_inner_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            cross_attention_dim=None,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = nn.LayerNorm(time_mix_inner_dim)
+            self.attn2 = Attention(
+                query_dim=time_mix_inner_dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        self.norm3 = nn.LayerNorm(time_mix_inner_dim)
+        self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu")
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = None
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        # chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off
+        self._chunk_dim = 1
+
+    def forward(
+        self,
+        hidden_states: paddle.Tensor,
+        num_frames: int,
+        encoder_hidden_states: Optional[paddle.Tensor] = None,
+    ) -> paddle.Tensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        batch_frames, seq_length, channels = hidden_states.shape
+        batch_size = batch_frames // num_frames
+
+        hidden_states = hidden_states[None, :].reshape([batch_size, num_frames, seq_length, channels])
+        hidden_states = hidden_states.transpose([0, 2, 1, 3])
+        hidden_states = hidden_states.reshape([batch_size * seq_length, num_frames, channels])
+
+        residual = hidden_states
+        hidden_states = self.norm_in(hidden_states)
+
+        if self._chunk_size is not None:
+            hidden_states = _chunked_feed_forward(self.ff, hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            hidden_states = self.ff_in(hidden_states)
+
+        if self.is_res:
+            hidden_states = hidden_states + residual
+
+        norm_hidden_states = self.norm1(hidden_states)
+        attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None)
+        hidden_states = attn_output + hidden_states
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            norm_hidden_states = self.norm2(hidden_states)
+            attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+
+        if self._chunk_size is not None:
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.is_res:
+            hidden_states = ff_output + hidden_states
+        else:
+            hidden_states = ff_output
+
+        hidden_states = hidden_states[None, :].reshape([batch_size, seq_length, num_frames, channels])
+        hidden_states = hidden_states.transpose([0, 2, 1, 3])
+        hidden_states = hidden_states.reshape([batch_size * num_frames, seq_length, channels])
+
+        return hidden_states
+
+
+class FeedForward(nn.Layer):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        dim (`int`): The number of channels in the input.
+        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
+        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        dim_out: Optional[int] = None,
+        mult: int = 4,
+        dropout: float = 0.0,
+        activation_fn: str = "geglu",
+        final_dropout: bool = False,
+        inner_dim=None,
+        bias: bool = True,
+    ):
+        super().__init__()
+        if inner_dim is None:
+            inner_dim = int(dim * mult)
+        dim_out = dim_out if dim_out is not None else dim
+        linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
+
+        if activation_fn == "gelu":
+            act_fn = GELU(dim, inner_dim, bias=bias)
+        if activation_fn == "gelu-approximate":
+            act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias)
+        elif activation_fn == "geglu":
+            act_fn = GEGLU(dim, inner_dim, bias=bias)
+        elif activation_fn == "geglu-approximate":
+            act_fn = ApproximateGELU(dim, inner_dim, bias=bias)
+
+        self.net = nn.LayerList([])
+        # project in
+        self.net.append(act_fn)
+        # project dropout
+        self.net.append(nn.Dropout(dropout))
+        # project out
+        self.net.append(linear_cls(inner_dim, dim_out, bias_attr=bias))
+        # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
+        if final_dropout:
+            self.net.append(nn.Dropout(dropout))
+
+    def forward(self, hidden_states: paddle.Tensor, scale: float = 1.0) -> paddle.Tensor:
+        compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear)
+        for module in self.net:
+            if isinstance(module, compatible_cls):
+                hidden_states = module(hidden_states, scale)
+            else:
+                hidden_states = module(hidden_states)
+        return hidden_states

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/models/autoencoder_tiny.py

@@ -0,0 +1,363 @@
+# Copyright 2023 Ollin Boer Bohan and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import paddle
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from ..utils.accelerate_utils import apply_forward_hook
+from .modeling_utils import ModelMixin
+from .vae import DecoderOutput, DecoderTiny, EncoderTiny
+
+
+@dataclass
+class AutoencoderTinyOutput(BaseOutput):
+    """
+    Output of AutoencoderTiny encoding method.
+
+    Args:
+        latents (`paddle.Tensor`): Encoded outputs of the `Encoder`.
+
+    """
+
+    latents: paddle.Tensor
+
+
+class AutoencoderTiny(ModelMixin, ConfigMixin):
+    r"""
+    A tiny distilled VAE model for encoding images into latents and decoding latent representations into images.
+
+    [`AutoencoderTiny`] is a wrapper around the original implementation of `TAESD`.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for its generic methods implemented for
+    all models (such as downloading or saving).
+
+    Parameters:
+        in_channels (`int`, *optional*, defaults to 3): Number of channels in the input image.
+        out_channels (`int`,  *optional*, defaults to 3): Number of channels in the output.
+        encoder_block_out_channels (`Tuple[int]`, *optional*, defaults to `(64, 64, 64, 64)`):
+            Tuple of integers representing the number of output channels for each encoder block. The length of the
+            tuple should be equal to the number of encoder blocks.
+        decoder_block_out_channels (`Tuple[int]`, *optional*, defaults to `(64, 64, 64, 64)`):
+            Tuple of integers representing the number of output channels for each decoder block. The length of the
+            tuple should be equal to the number of decoder blocks.
+        act_fn (`str`, *optional*, defaults to `"relu"`):
+            Activation function to be used throughout the model.
+        latent_channels (`int`, *optional*, defaults to 4):
+            Number of channels in the latent representation. The latent space acts as a compressed representation of
+            the input image.
+        upsampling_scaling_factor (`int`, *optional*, defaults to 2):
+            Scaling factor for upsampling in the decoder. It determines the size of the output image during the
+            upsampling process.
+        num_encoder_blocks (`Tuple[int]`, *optional*, defaults to `(1, 3, 3, 3)`):
+            Tuple of integers representing the number of encoder blocks at each stage of the encoding process. The
+            length of the tuple should be equal to the number of stages in the encoder. Each stage has a different
+            number of encoder blocks.
+        num_decoder_blocks (`Tuple[int]`, *optional*, defaults to `(3, 3, 3, 1)`):
+            Tuple of integers representing the number of decoder blocks at each stage of the decoding process. The
+            length of the tuple should be equal to the number of stages in the decoder. Each stage has a different
+            number of decoder blocks.
+        latent_magnitude (`float`, *optional*, defaults to 3.0):
+            Magnitude of the latent representation. This parameter scales the latent representation values to control
+            the extent of information preservation.
+        latent_shift (float, *optional*, defaults to 0.5):
+            Shift applied to the latent representation. This parameter controls the center of the latent space.
+        scaling_factor (`float`, *optional*, defaults to 1.0):
+            The component-wise standard deviation of the trained latent space computed using the first batch of the
+            training set. This is used to scale the latent space to have unit variance when training the diffusion
+            model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+            diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+            / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+            Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper. For this Autoencoder,
+            however, no such scaling factor was used, hence the value of 1.0 as the default.
+        force_upcast (`bool`, *optional*, default to `False`):
+            If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
+            can be fine-tuned / trained to a lower range without losing too much precision, in which case
+            `force_upcast` can be set to `False` (see this fp16-friendly
+            [AutoEncoder](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix)).
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        encoder_block_out_channels: Tuple[int, ...] = (64, 64, 64, 64),
+        decoder_block_out_channels: Tuple[int, ...] = (64, 64, 64, 64),
+        act_fn: str = "relu",
+        latent_channels: int = 4,
+        upsampling_scaling_factor: int = 2,
+        num_encoder_blocks: Tuple[int, ...] = (1, 3, 3, 3),
+        num_decoder_blocks: Tuple[int, ...] = (3, 3, 3, 1),
+        latent_magnitude: int = 3,
+        latent_shift: float = 0.5,
+        force_upcast: bool = False,
+        scaling_factor: float = 1.0,
+    ):
+        super().__init__()
+
+        if len(encoder_block_out_channels) != len(num_encoder_blocks):
+            raise ValueError("`encoder_block_out_channels` should have the same length as `num_encoder_blocks`.")
+        if len(decoder_block_out_channels) != len(num_decoder_blocks):
+            raise ValueError("`decoder_block_out_channels` should have the same length as `num_decoder_blocks`.")
+
+        self.encoder = EncoderTiny(
+            in_channels=in_channels,
+            out_channels=latent_channels,
+            num_blocks=num_encoder_blocks,
+            block_out_channels=encoder_block_out_channels,
+            act_fn=act_fn,
+        )
+
+        self.decoder = DecoderTiny(
+            in_channels=latent_channels,
+            out_channels=out_channels,
+            num_blocks=num_decoder_blocks,
+            block_out_channels=decoder_block_out_channels,
+            upsampling_scaling_factor=upsampling_scaling_factor,
+            act_fn=act_fn,
+        )
+
+        self.latent_magnitude = latent_magnitude
+        self.latent_shift = latent_shift
+        self.scaling_factor = scaling_factor
+
+        self.use_slicing = False
+        self.use_tiling = False
+
+        # only relevant if vae tiling is enabled
+        self.spatial_scale_factor = 2**out_channels
+        self.tile_overlap_factor = 0.125
+        self.tile_sample_min_size = 512
+        self.tile_latent_min_size = self.tile_sample_min_size // self.spatial_scale_factor
+
+        self.register_to_config(block_out_channels=decoder_block_out_channels)
+        self.register_to_config(force_upcast=False)
+
+    def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+        if isinstance(module, (EncoderTiny, DecoderTiny)):
+            module.gradient_checkpointing = value
+
+    def scale_latents(self, x: paddle.Tensor) -> paddle.Tensor:
+        """raw latents -> [0, 1]"""
+        return ((x / 2 * self.latent_magnitude) + self.latent_shift).clip(0, 1)
+
+    def unscale_latents(self, x: paddle.Tensor) -> paddle.Tensor:
+        """[0, 1] -> raw latents"""
+        return (x - self.latent_shift) * (2 * self.latent_magnitude)
+
+    def enable_slicing(self) -> None:
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.use_slicing = True
+
+    def disable_slicing(self) -> None:
+        r"""
+        Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.use_slicing = False
+
+    def enable_tiling(self, use_tiling: bool = True) -> None:
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.use_tiling = use_tiling
+
+    def disable_tiling(self) -> None:
+        r"""
+        Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.enable_tiling(False)
+
+    def _tiled_encode(self, x: paddle.Tensor) -> paddle.Tensor:
+        r"""Encode a batch of images using a tiled encoder.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
+        steps. This is useful to keep memory use constant regardless of image size. To avoid tiling artifacts, the
+        tiles overlap and are blended together to form a smooth output.
+
+        Args:
+            x (`paddle.Tensor`): Input batch of images.
+
+        Returns:
+            `paddle.Tensor`: Encoded batch of images.
+        """
+        # scale of encoder output relative to input
+        sf = self.spatial_scale_factor
+        tile_size = self.tile_sample_min_size
+
+        # number of pixels to blend and to traverse between tile
+        blend_size = int(tile_size * self.tile_overlap_factor)
+        traverse_size = tile_size - blend_size
+
+        # tiles index (up/left)
+        ti = range(0, x.shape[-2], traverse_size)
+        tj = range(0, x.shape[-1], traverse_size)
+
+        # mask for blending
+        blend_masks = paddle.stack(
+            paddle.meshgrid([paddle.arange(tile_size / sf) / (blend_size / sf - 1)] * 2, indexing="ij")
+        )
+        blend_masks = blend_masks.clip(0, 1)
+
+        # output array
+        out = paddle.zeros([x.shape[0], 4, x.shape[-2] // sf, x.shape[-1] // sf])
+        for i in ti:
+            for j in tj:
+                tile_in = x[..., i : i + tile_size, j : j + tile_size]
+                # tile result
+                tile_out = out[..., i // sf : (i + tile_size) // sf, j // sf : (j + tile_size) // sf]
+                tile = self.encoder(tile_in)
+                h, w = tile.shape[-2], tile.shape[-1]
+                # blend tile result into output
+                blend_mask_i = paddle.ones_like(blend_masks[0]) if i == 0 else blend_masks[0]
+                blend_mask_j = paddle.ones_like(blend_masks[1]) if j == 0 else blend_masks[1]
+                blend_mask = blend_mask_i * blend_mask_j
+                tile, blend_mask = tile[..., :h, :w], blend_mask[..., :h, :w]
+
+                # NOTE this copy_ method is not work in paddlepaddle
+                # tile_out.copy_(blend_mask * tile + (1 - blend_mask) * tile_out)
+                out[..., i // sf : (i + tile_size) // sf, j // sf : (j + tile_size) // sf] = (
+                    blend_mask * tile + (1 - blend_mask) * tile_out
+                )
+
+        return out
+
+    def _tiled_decode(self, x: paddle.Tensor) -> paddle.Tensor:
+        r"""Encode a batch of images using a tiled encoder.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
+        steps. This is useful to keep memory use constant regardless of image size. To avoid tiling artifacts, the
+        tiles overlap and are blended together to form a smooth output.
+
+        Args:
+            x (`paddle.Tensor`): Input batch of images.
+
+        Returns:
+            `paddle.Tensor`: Encoded batch of images.
+        """
+        # scale of decoder output relative to input
+        sf = self.spatial_scale_factor
+        tile_size = self.tile_latent_min_size
+
+        # number of pixels to blend and to traverse between tiles
+        blend_size = int(tile_size * self.tile_overlap_factor)
+        traverse_size = tile_size - blend_size
+
+        # tiles index (up/left)
+        ti = range(0, x.shape[-2], traverse_size)
+        tj = range(0, x.shape[-1], traverse_size)
+
+        # mask for blending
+        blend_masks = paddle.stack(
+            paddle.meshgrid([paddle.arange(tile_size * sf) / (blend_size * sf - 1)] * 2, indexing="ij")
+        )
+        blend_masks = blend_masks.clip(0, 1)
+
+        # output array
+        out = paddle.zeros([x.shape[0], 3, x.shape[-2] * sf, x.shape[-1] * sf])
+        for i in ti:
+            for j in tj:
+                tile_in = x[..., i : i + tile_size, j : j + tile_size]
+                # tile result
+                tile_out = out[..., i * sf : (i + tile_size) * sf, j * sf : (j + tile_size) * sf]
+                tile = self.decoder(tile_in)
+                h, w = tile.shape[-2], tile.shape[-1]
+                # blend tile result into output
+                blend_mask_i = paddle.ones_like(blend_masks[0]) if i == 0 else blend_masks[0]
+                blend_mask_j = paddle.ones_like(blend_masks[1]) if j == 0 else blend_masks[1]
+                blend_mask = (blend_mask_i * blend_mask_j)[..., :h, :w]
+
+                # NOTE this copy_ method is not work in paddlepaddle
+                # tile_out.copy_(blend_mask * tile + (1 - blend_mask) * tile_out)
+                out[..., i * sf : (i + tile_size) * sf, j * sf : (j + tile_size) * sf] = (
+                    blend_mask * tile + (1 - blend_mask) * tile_out
+                )
+        return out
+
+    @apply_forward_hook
+    def encode(self, x: paddle.Tensor, return_dict: bool = True) -> Union[AutoencoderTinyOutput, Tuple[paddle.Tensor]]:
+        # TODO junnyu, support float16
+        x = x.cast(self.encoder.layers[0].weight.dtype)
+        if self.use_slicing and x.shape[0] > 1:
+            output = [
+                self._tiled_encode(x_slice) if self.use_tiling else self.encoder(x) for x_slice in x.chunk(x.shape[0])
+            ]
+            output = paddle.concat(output)
+        else:
+            output = self._tiled_encode(x) if self.use_tiling else self.encoder(x)
+
+        if not return_dict:
+            return (output,)
+
+        return AutoencoderTinyOutput(latents=output)
+
+    @apply_forward_hook
+    def decode(
+        self, x: paddle.Tensor, generator: Optional[paddle.Generator] = None, return_dict: bool = True
+    ) -> Union[DecoderOutput, Tuple[paddle.Tensor]]:
+        # TODO junnyu, add this to support pure fp16
+        x = x.cast(self.decoder.layers[0].weight.dtype)
+
+        if self.use_slicing and x.shape[0] > 1:
+            output = [
+                self._tiled_decode(x_slice) if self.use_tiling else self.decoder(x) for x_slice in x.chunk(x.shape[0])
+            ]
+            output = paddle.concat(output)
+        else:
+            output = self._tiled_decode(x) if self.use_tiling else self.decoder(x)
+
+        if not return_dict:
+            return (output,)
+
+        return DecoderOutput(sample=output)
+
+    def forward(
+        self,
+        sample: paddle.Tensor,
+        return_dict: bool = True,
+    ) -> Union[DecoderOutput, Tuple[paddle.Tensor]]:
+        r"""
+        Args:
+            sample (`paddle.Tensor`): Input sample.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+        """
+        enc = self.encode(sample).latents
+
+        # scale latents to be in [0, 1], then quantize latents to a byte tensor,
+        # as if we were storing the latents in an RGBA uint8 image.
+        scaled_enc = (self.scale_latents(enc) * 255).round().cast("byte")
+
+        # unquantize latents back into [0, 1], then unscale latents back to their original range,
+        # as if we were loading the latents from an RGBA uint8 image.
+        unscaled_enc = self.unscale_latents(scaled_enc / 255.0)
+
+        dec = self.decode(unscaled_enc)
+
+        if not return_dict:
+            return (dec,)
+        return DecoderOutput(sample=dec)

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/models/controlnet_sd3.py

@@ -0,0 +1,422 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+from paddle.distributed.fleet.utils import recompute
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..loaders import FromOriginalControlnetMixin
+from ..models.attention import JointTransformerBlock
+from ..models.attention_processor import Attention, AttentionProcessor
+from ..models.modeling_outputs  import Transformer2DModelOutput
+from ..models.modeling_utils import ModelMixin
+from ..utils import (
+    USE_PEFT_BACKEND,
+    logging,
+    recompute_use_reentrant,
+    scale_lora_layers,
+    unscale_lora_layers,
+    use_old_recompute,
+)
+from .controlnet import BaseOutput, zero_module
+from .embeddings import CombinedTimestepTextProjEmbeddings, PatchEmbed
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class SD3ControlNetOutput(BaseOutput):
+    controlnet_block_samples: Tuple[paddle.Tensor]
+
+
+class SD3ControlNetModel(ModelMixin, ConfigMixin, FromOriginalControlnetMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: int = 128,
+        patch_size: int = 2,
+        in_channels: int = 16,
+        num_layers: int = 18,
+        attention_head_dim: int = 64,
+        num_attention_heads: int = 18,
+        joint_attention_dim: int = 4096,
+        caption_projection_dim: int = 1152,
+        pooled_projection_dim: int = 2048,
+        out_channels: int = 16,
+        pos_embed_max_size: int = 96,
+        extra_conditioning_channels: int = 0,
+    ):
+        super().__init__()
+        default_out_channels = in_channels
+        self.out_channels = out_channels if out_channels is not None else default_out_channels
+        self.inner_dim = num_attention_heads * attention_head_dim
+
+        self.pos_embed = PatchEmbed(
+            height=sample_size,
+            width=sample_size,
+            patch_size=patch_size,
+            in_channels=in_channels,
+            embed_dim=self.inner_dim,
+            pos_embed_max_size=pos_embed_max_size,
+        )
+        self.time_text_embed = CombinedTimestepTextProjEmbeddings(
+            embedding_dim=self.inner_dim, pooled_projection_dim=pooled_projection_dim
+        )
+        self.context_embedder = nn.Linear(joint_attention_dim, caption_projection_dim)
+
+        # `attention_head_dim` is doubled to account for the mixing.
+        # It needs to crafted when we get the actual checkpoints.
+        self.transformer_blocks = nn.LayerList(
+            [
+                JointTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    # attention_head_dim=self.config.attention_head_dim,
+                    attention_head_dim=self.inner_dim,
+                    context_pre_only=False,
+                )
+                for i in range(num_layers)
+            ]
+        )
+
+        # controlnet_blocks
+        self.controlnet_blocks = nn.LayerList([])
+        for _ in range(len(self.transformer_blocks)):
+            controlnet_block = nn.Linear(self.inner_dim, self.inner_dim)
+            controlnet_block = zero_module(controlnet_block)
+            self.controlnet_blocks.append(controlnet_block)
+        pos_embed_input = PatchEmbed(
+            height=sample_size,
+            width=sample_size,
+            patch_size=patch_size,
+            in_channels=in_channels + extra_conditioning_channels,
+            embed_dim=self.inner_dim,
+            add_pos_embed=False, # to verify
+        )
+        self.pos_embed_input = zero_module(pos_embed_input)
+
+        self.gradient_checkpointing = False
+
+    # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.enable_forward_chunking
+    def enable_forward_chunking(self, chunk_size: Optional[int] = None, dim: int = 0) -> None:
+        """
+        Sets the attention processor to use [feed forward
+        chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).
+        Parameters:
+            chunk_size (`int`, *optional*):
+                The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
+                over each tensor of dim=`dim`.
+            dim (`int`, *optional*, defaults to `0`):
+                The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
+                or dim=1 (sequence length).
+        """
+        if dim not in [0, 1]:
+            raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")
+
+        # By default chunk size is 1
+        chunk_size = chunk_size or 1
+
+        def fn_recursive_feed_forward(module: paddle.nn.Layer, chunk_size: int, dim: int):
+            if hasattr(module, "set_chunk_feed_forward"):
+                module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+            for child in module.children():
+                fn_recursive_feed_forward(child, chunk_size, dim)
+
+        for module in self.children():
+            fn_recursive_feed_forward(module, chunk_size, dim)
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: paddle.nn.Layer, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: paddle.nn.Layer, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
+        are fused. For cross-attention modules, key and value projection matrices are fused.
+        <Tip warning={true}>
+        This API is 🧪 experimental.
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+        
+        # TODO: ?
+        # self.set_attn_processor(FusedJointAttnProcessor2_0())
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+        <Tip warning={true}>
+        This API is 🧪 experimental.
+        </Tip>
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+     
+    #  TODO
+    @classmethod
+    def from_transformer(
+        cls, transformer, num_layers=12, num_extra_conditioning_channels=1, load_weights_from_transformer=True
+    ):
+        config = transformer.config
+        config["num_layers"] = num_layers or config.num_layers
+        config["extra_conditioning_channels"] = num_extra_conditioning_channels
+        controlnet = cls(**config)
+
+        if load_weights_from_transformer:
+            controlnet.pos_embed.load_dict(transformer.pos_embed.state_dict())
+            controlnet.time_text_embed.load_dict(transformer.time_text_embed.state_dict())
+            controlnet.context_embedder.load_dict(transformer.context_embedder.state_dict())
+            controlnet.transformer_blocks.load_dict(transformer.transformer_blocks.state_dict(), strict=False)
+
+            controlnet.pos_embed_input = zero_module(controlnet.pos_embed_input)
+
+        return controlnet
+
+    def forward(
+        self,
+        hidden_states: paddle.Tensor,
+        controlnet_cond: paddle.Tensor,
+        conditioning_scale: float = 1.0,
+        encoder_hidden_states: paddle.Tensor = None,
+        pooled_projections: paddle.Tensor = None,
+        timestep: paddle.Tensor = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[paddle.Tensor, Transformer2DModelOutput]:
+        """
+        The [`SD3Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`paddle.Tensor` of shape `(batch size, channel, height, width)`):
+                Input `hidden_states`.
+            controlnet_cond (`paddle.Tensor`):
+                The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
+            conditioning_scale (`float`, defaults to `1.0`):
+                The scale factor for ControlNet outputs.
+            encoder_hidden_states (`paddle.Tensor` of shape `(batch size, sequence_len, embed_dims)`):
+                Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
+            pooled_projections (`paddle.Tensor` of shape `(batch_size, projection_dim)`): Embeddings projected
+                from the embeddings of input conditions.
+            timestep ( `paddle.Tensor`):
+                Used to indicate denoising step.
+            joint_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        if joint_attention_kwargs is not None:
+            joint_attention_kwargs = joint_attention_kwargs.copy()
+            lora_scale = joint_attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        hidden_states = self.pos_embed(hidden_states)  # takes care of adding positional embeddings too.
+        temb = self.time_text_embed(timestep, pooled_projections)
+        encoder_hidden_states = self.context_embedder(encoder_hidden_states)
+
+        # add
+        hidden_states = hidden_states + self.pos_embed_input(controlnet_cond)
+
+        block_res_samples = ()
+
+        for block in self.transformer_blocks:
+            if self.training and self.gradient_checkpointing and not use_old_recompute():
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs = {} if recompute_use_reentrant() else {"use_reentrant": False}
+                hidden_states = recompute(
+                    create_custom_forward(block),
+                    hidden_states,
+                    encoder_hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+
+            else:
+                encoder_hidden_states, hidden_states = block(
+                    hidden_states=hidden_states, encoder_hidden_states=encoder_hidden_states, temb=temb
+                )
+
+            block_res_samples = block_res_samples + (hidden_states,)
+
+        controlnet_block_res_samples = ()
+        for block_res_sample, controlnet_block in zip(block_res_samples, self.controlnet_blocks):
+            block_res_sample = controlnet_block(block_res_sample)
+            controlnet_block_res_samples = controlnet_block_res_samples + (block_res_sample,)
+
+        # 6. scaling
+        controlnet_block_res_samples = [sample * conditioning_scale for sample in controlnet_block_res_samples]
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (controlnet_block_res_samples,)
+
+        return SD3ControlNetOutput(controlnet_block_samples=controlnet_block_res_samples)
+
+
+class SD3MultiControlNetModel(ModelMixin):
+    r"""
+    `SD3ControlNetModel` wrapper class for Multi-SD3ControlNet
+
+    This module is a wrapper for multiple instances of the `SD3ControlNetModel`. The `forward()` API is designed to be
+    compatible with `SD3ControlNetModel`.
+
+    Args:
+        controlnets (`List[SD3ControlNetModel]`):
+            Provides additional conditioning to the unet during the denoising process. You must set multiple
+            `SD3ControlNetModel` as a list.
+    """
+
+    def __init__(self, controlnets):
+        super().__init__()
+        self.nets = nn.LayerList(controlnets)
+
+    def forward(
+        self,
+        hidden_states: paddle.Tensor,
+        controlnet_cond: List[paddle.Tensor],
+        conditioning_scale: List[float],
+        pooled_projections: paddle.Tensor,
+        encoder_hidden_states: paddle.Tensor = None,
+        timestep: paddle.Tensor = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[SD3ControlNetOutput, Tuple]:
+        for i, (image, scale, controlnet) in enumerate(zip(controlnet_cond, conditioning_scale, self.nets)):
+            block_samples = controlnet(
+                hidden_states=hidden_states,
+                timestep=timestep,
+                encoder_hidden_states=encoder_hidden_states,
+                pooled_projections=pooled_projections,
+                controlnet_cond=image,
+                conditioning_scale=scale,
+                joint_attention_kwargs=joint_attention_kwargs,
+                return_dict=return_dict,
+            )
+
+            # merge samples
+            if i == 0:
+                control_block_samples = block_samples
+            else:
+                control_block_samples = [
+                    control_block_sample + block_sample
+                    for control_block_sample, block_sample in zip(control_block_samples[0], block_samples[0])
+                ]
+                control_block_samples = (tuple(control_block_samples),)
+
+        return control_block_samples

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/models/dit_llama.py

@@ -0,0 +1,576 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Optional
+
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+from paddle.nn.functional.flash_attention import flash_attention
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from .embeddings import LabelEmbedding
+from .modeling_utils import ModelMixin
+from .transformer_2d import Transformer2DModelOutput
+
+
+def TypePromote(x, y):
+    TYPE_PROMOTE_DICT = {
+        "INT16FP16": "float16",
+        "INT16FP32": "float32",
+        "INT16FP64": "float64",
+        "INT32FP16": "float32",
+        "INT32FP32": "float32",
+        "INT32FP64": "float64",
+        "INT64FP16": "float64",
+        "INT64FP32": "float64",
+        "INT64FP64": "float64",
+    }
+    if x.dtype.name + y.dtype.name in TYPE_PROMOTE_DICT:
+        promote_type = TYPE_PROMOTE_DICT[x.dtype.name + y.dtype.name]
+    elif y.dtype.name + x.dtype.name in TYPE_PROMOTE_DICT:
+        promote_type = TYPE_PROMOTE_DICT[y.dtype.name + x.dtype.name]
+    else:
+        return x, y
+    return x.cast(promote_type), y.cast(promote_type)
+
+
+def modulate(x, shift, scale):
+    return x * (1 + scale.unsqueeze(axis=1)) + shift.unsqueeze(axis=1)
+
+
+class TimestepEmbedder(nn.Layer):
+    """
+    Embeds scalar timesteps into vector representations.
+    """
+
+    def __init__(self, hidden_size, frequency_embedding_size=256):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            nn.Linear(frequency_embedding_size, hidden_size),
+            nn.Silu(),
+            nn.Linear(hidden_size, hidden_size),
+        )
+        self.frequency_embedding_size = frequency_embedding_size
+
+    @staticmethod
+    def timestep_embedding(t, dim, max_period=10000):
+        """
+        Create sinusoidal timestep embeddings.
+        :param t: a 1-D Tensor of N indices, one per batch element.
+                          These may be fractional.
+        :param dim: the dimension of the output.
+        :param max_period: controls the minimum frequency of the embeddings.
+        :return: an (N, D) Tensor of positional embeddings.
+        """
+        half = dim // 2
+        freqs = paddle.exp(x=-math.log(max_period) * paddle.arange(start=0, end=half, dtype="float32") / half)
+        args = t[:, (None)].astype(dtype="float32") * freqs[None]
+        embedding = paddle.concat(x=[paddle.cos(x=args), paddle.sin(x=args)], axis=-1)
+        if dim % 2:
+            embedding = paddle.concat(x=[embedding, paddle.zeros_like(x=embedding[:, :1])], axis=-1)
+        return embedding
+
+    def forward(self, t):
+        t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
+        t_emb = self.mlp(t_freq.cast(self.mlp[0].weight.dtype))
+        return t_emb
+
+
+class Attention(nn.Layer):
+    def __init__(self, dim, n_heads, n_kv_heads, qk_norm=True, fused_attn=True):
+        """
+        Initialize the Attention module.
+
+        Args:
+            dim (int): Number of input dimensions.
+            n_heads (int): Number of heads.
+            n_kv_heads (Optional[int]): Number of kv heads, if using GQA.
+
+        Attributes:
+            n_kv_heads (int): Number of key and value heads.
+            n_local_heads (int): Number of local query heads.
+            n_local_kv_heads (int): Number of local key and value heads.
+            n_rep (int): Number of repetitions for local heads.
+            head_dim (int): Dimension size of each attention head.
+            wq (nn.Linear): Linear transformation for queries.
+            wk (nn.Linear): Linear transformation for keys.
+            wv (nn.Linear): Linear transformation for values.
+            wo (nn.Linear): Linear transformation for output.
+            cache_k (paddle.Tensor): Cached keys for attention.
+            cache_v (paddle.Tensor): Cached values for attention.
+
+        """
+        super().__init__()
+        self.n_kv_heads = n_heads if n_kv_heads is None else n_kv_heads
+        self.n_local_heads = n_heads
+        self.n_local_kv_heads = self.n_kv_heads
+        self.n_rep = self.n_local_heads // self.n_local_kv_heads
+        self.head_dim = dim // n_heads
+
+        self.wq = nn.Linear(dim, n_heads * self.head_dim, bias_attr=False)
+        self.wk = nn.Linear(dim, self.n_kv_heads * self.head_dim, bias_attr=False)
+        self.wv = nn.Linear(dim, self.n_kv_heads * self.head_dim, bias_attr=False)
+        self.wo = nn.Linear(n_heads * self.head_dim, dim, bias_attr=False)
+
+        if qk_norm:
+            self.q_norm = nn.LayerNorm(self.n_local_heads * self.head_dim)
+            self.k_norm = nn.LayerNorm(self.n_local_kv_heads * self.head_dim)
+        else:
+            self.q_norm = self.k_norm = nn.Identity()
+
+        self.fused_attn = fused_attn
+        self.scale = self.head_dim**-0.5
+
+    @staticmethod
+    def reshape_for_broadcast(freqs_cis, x):
+        """
+        Reshape frequency tensor for broadcasting it with another tensor.
+
+        This function reshapes the frequency tensor to have the same shape as
+        the target tensor 'x' for the purpose of broadcasting the frequency
+        tensor during element-wise operations.
+
+        Args:
+            freqs_cis (paddle.Tensor): Frequency tensor to be reshaped.
+            x (paddle.Tensor): Target tensor for broadcasting compatibility.
+
+        Returns:
+            paddle.Tensor: Reshaped frequency tensor.
+
+        Raises:
+            AssertionError: If the frequency tensor doesn't match the expected
+                shape.
+            AssertionError: If the target tensor 'x' doesn't have the expected
+                number of dimensions.
+        """
+        ndim = x.ndim
+        assert 0 <= 1 < ndim
+        assert tuple(freqs_cis.shape) == (tuple(x.shape)[1], tuple(x.shape)[-1])
+        shape = [(d if i == 1 or i == ndim - 1 else 1) for i, d in enumerate(tuple(x.shape))]
+        return freqs_cis.reshape([*shape])
+
+    @staticmethod
+    def apply_rotary_emb(xq, xk, freqs_cis):
+        """
+        Apply rotary embeddings to input tensors using the given frequency
+        tensor.
+
+        This function applies rotary embeddings to the given query 'xq' and
+        key 'xk' tensors using the provided frequency tensor 'freqs_cis'. The
+        input tensors are reshaped as complex numbers, and the frequency tensor
+        is reshaped for broadcasting compatibility. The resulting tensors
+        contain rotary embeddings and are returned as real tensors.
+
+        Args:
+            xq (paddle.Tensor): Query tensor to apply rotary embeddings.
+            xk (paddle.Tensor): Key tensor to apply rotary embeddings.
+            freqs_cis (paddle.Tensor): Precomputed frequency tensor for complex
+                exponentials.
+
+        Returns:
+            Tuple[paddle.Tensor, paddle.Tensor]: Tuple of modified query tensor
+                and key tensor with rotary embeddings.
+        """
+        with paddle.amp.auto_cast(enable=False):
+            xq_ = paddle.as_complex(xq.cast("float32").reshape([*tuple(xq.shape)[:-1], -1, 2]))
+            xk_ = paddle.as_complex(xk.cast("float32").reshape([*tuple(xk.shape)[:-1], -1, 2]))
+            freqs_cis = Attention.reshape_for_broadcast(freqs_cis, xq_)
+            xq_out = paddle.as_real(xq_ * freqs_cis).flatten(start_axis=3)
+            xk_out = paddle.as_real(xk_ * freqs_cis).flatten(start_axis=3)
+            return xq_out.cast(xq.dtype), xk_out.cast(xk.dtype)
+
+    def forward(self, x, freqs_cis):
+        """
+        Forward pass of the attention module.
+
+        Args:
+            x (paddle.Tensor): Input tensor.
+            freqs_cis (paddle.Tensor): Precomputed frequency tensor.
+
+        Returns:
+            paddle.Tensor: Output tensor after attention.
+
+        """
+        bsz, seqlen, _ = tuple(x.shape)
+        xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)
+        dtype = xq.dtype
+
+        xq = self.q_norm(xq)
+        xk = self.k_norm(xk)
+
+        xq = xq.reshape([bsz, seqlen, self.n_local_heads, self.head_dim])
+        xk = xk.reshape([bsz, seqlen, self.n_local_kv_heads, self.head_dim])
+        xv = xv.reshape([bsz, seqlen, self.n_local_kv_heads, self.head_dim])
+
+        xq, xk = Attention.apply_rotary_emb(xq, xk, freqs_cis=freqs_cis)
+        xq, xk = xq.cast(dtype), xk.cast(dtype)
+
+        if dtype in [paddle.float16, paddle.bfloat16]:
+            output, _ = flash_attention(
+                xq,
+                xk,
+                xv,
+                dropout=0.0,
+                causal=False,
+                return_softmax=False,
+            )
+        else:
+            n_rep = self.n_local_heads // self.n_local_kv_heads
+            if n_rep > 1:
+                xk = xk.unsqueeze(axis=3).tile([1, 1, 1, n_rep, 1]).flatten(start_axis=2, stop_axis=3)
+                xv = xv.unsqueeze(axis=3).tile([1, 1, 1, n_rep, 1]).flatten(start_axis=2, stop_axis=3)
+
+            if self.fused_attn:
+                output = F.scaled_dot_product_attention_(
+                    xq,
+                    xk,
+                    xv,
+                    dropout_p=0.0,
+                    is_causal=False,
+                )
+            else:
+                q = xq.transpose([0, 2, 1, 3]) * self.scale
+                attn = q @ xk.transpose([0, 2, 1, 3]).transpose([0, 1, 3, 2])
+                attn = F.softmax(attn, axis=-1)
+                output = attn @ xv.transpose([0, 2, 1, 3])
+                output = output.transpose([0, 2, 1, 3])
+
+        output = output.flatten(start_axis=-2)
+        return self.wo(output)
+
+
+class FeedForward(nn.Layer):
+    def __init__(self, dim, hidden_dim, multiple_of=256, ffn_dim_multiplier=None):
+        """
+        Initialize the FeedForward module.
+
+        Args:
+            dim (int): Input dimension.
+            hidden_dim (int): Hidden dimension of the feedforward layer.
+            multiple_of (int): Value to ensure hidden dimension is a multiple
+                of this value.
+            ffn_dim_multiplier (float, optional): Custom multiplier for hidden
+                dimension. Defaults to None.
+
+        Attributes:
+            w1 (nn.Linear): Linear transformation for the first
+                layer.
+            w2 (nn.Linear): Linear transformation for the second layer.
+            w3 (nn.Linear): Linear transformation for the third
+                layer.
+
+        """
+        super().__init__()
+        hidden_dim = int(2 * hidden_dim / 3)
+        if ffn_dim_multiplier is not None:
+            hidden_dim = int(ffn_dim_multiplier * hidden_dim)
+        hidden_dim = int(multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of))
+
+        self.w1 = nn.Linear(dim, hidden_dim, bias_attr=False)
+        self.w2 = nn.Linear(hidden_dim, dim, bias_attr=False)
+        self.w3 = nn.Linear(dim, hidden_dim, bias_attr=False)
+
+    def forward(self, x):
+        xw1 = F.silu(self.w1(x))
+        xw3 = self.w3(x)
+        output = self.w2(xw1 * xw3)
+        return output
+
+
+class TransformerBlock(nn.Layer):
+    def __init__(
+        self,
+        layer_id: int,
+        dim: int,
+        n_heads: int,
+        n_kv_heads: int,
+        multiple_of: int,
+        mlp_ratio: float,
+        ffn_dim_multiplier: float,
+        norm_eps: float,
+        qk_norm: bool,
+        fused_attn: bool,
+    ) -> None:
+        """
+        Initialize a TransformerBlock.
+
+        Args:
+            layer_id (int): Identifier for the layer.
+            dim (int): Embedding dimension of the input features.
+            n_heads (int): Number of attention heads.
+            n_kv_heads (Optional[int]): Number of attention heads in key and
+                value features (if using GQA), or set to None for the same as
+                query.
+            multiple_of (int): Value to ensure hidden dimension is a multiple
+                of this value in the FeedForward block.
+            ffn_dim_multiplier (float, optional): Custom multiplier for hidden
+                dimension in the FeedForward block. Defaults to None.
+            norm_eps (float): A small value added to the norm layer
+                denominators to avoid division-by-zero.
+
+        Attributes:
+            n_heads (int): Number of attention heads.
+            dim (int): Dimension size of the model.
+            head_dim (int): Dimension size of each attention head.
+            attention (Attention): Attention module.
+            feed_forward (FeedForward): FeedForward module.
+            layer_id (int): Identifier for the layer.
+            attention_norm (RMSNorm): Layer normalization for attention output.
+            ffn_norm (RMSNorm): Layer normalization for feedforward output.
+            adaLN_modulation (nn.Sequential): A small network to generate
+                feature modulation factors.
+
+        """
+        super().__init__()
+        self.dim = dim
+        self.head_dim = dim // n_heads
+        self.attention = Attention(dim, n_heads, n_kv_heads, qk_norm, fused_attn)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.feed_forward = FeedForward(
+            dim=dim, hidden_dim=mlp_hidden_dim, multiple_of=multiple_of, ffn_dim_multiplier=ffn_dim_multiplier
+        )
+        self.layer_id = layer_id
+        self.attention_norm = nn.LayerNorm(dim, epsilon=norm_eps, bias_attr=False)
+        self.ffn_norm = nn.LayerNorm(dim, epsilon=norm_eps, bias_attr=False)
+
+        self.adaLN_modulation = nn.Sequential(
+            nn.Silu(),
+            nn.Linear(min(dim, 1024), 6 * dim),
+        )
+
+    def forward(self, x, freqs_cis, adaln_input=None):
+        """
+        Perform a forward pass through the TransformerBlock.
+
+        Args:
+            x (paddle.Tensor): Input tensor.
+            freqs_cis (paddle.Tensor): Precomputed cosine and sine frequencies.
+            mask (paddle.Tensor, optional): Masking tensor for attention.
+                Defaults to None.
+
+        Returns:
+            paddle.Tensor: Output tensor after applying attention and
+                feedforward layers.
+
+        """
+        if adaln_input is not None:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).chunk(
+                6, axis=1
+            )
+            h = x + gate_msa.unsqueeze(1) * self.attention(
+                modulate(self.attention_norm(x), shift_msa, scale_msa), freqs_cis
+            )
+            out = h + gate_mlp.unsqueeze(1) * self.feed_forward(modulate(self.ffn_norm(h), shift_mlp, scale_mlp))
+        else:
+            h = x + self.attention(self.attention_norm(x), freqs_cis)
+            out = h + self.feed_forward(self.ffn_norm(h))
+        return out
+
+
+class FinalLayer(paddle.nn.Layer):
+    def __init__(self, hidden_size, patch_size, out_channels):
+        super().__init__()
+        self.norm_final = paddle.nn.LayerNorm(hidden_size, weight_attr=False, bias_attr=False, epsilon=1e-06)
+        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels)
+        self.adaLN_modulation = nn.Sequential(nn.Silu(), nn.Linear(min(hidden_size, 1024), 2 * hidden_size))
+
+    def forward(self, x, c):
+        shift, scale = self.adaLN_modulation(c).chunk(2, axis=1)
+        x = modulate(self.norm_final(x), shift, scale)
+        x = self.linear(x)
+        return x
+
+
+class DiTLLaMA2DModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+    _use_memory_efficient_attention_xformers = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: int = 32,  # image_size // 8
+        patch_size: int = 2,
+        in_channels: int = 4,
+        out_channels: int = 8,
+        num_layers: int = 32,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 96,
+        mlp_ratio: float = 4.0,
+        n_kv_heads=None,
+        multiple_of: int = 256,
+        ffn_dim_multiplier=None,
+        norm_eps: float = 1e-05,
+        class_dropout_prob: float = 0.1,
+        num_classes: int = 1000,
+        learn_sigma: bool = True,
+        qk_norm: bool = True,
+    ):
+        super().__init__()
+        self.sample_size = sample_size
+        self.patch_size = patch_size
+        self.in_channels = in_channels
+        self.out_channels = in_channels * 2 if learn_sigma else in_channels
+        dim = attention_head_dim * num_attention_heads
+
+        self.num_layers = num_layers
+        self.num_attention_heads = num_attention_heads
+        self.mlp_ratio = mlp_ratio
+        self.multiple_of = multiple_of
+        self.ffn_dim_multiplier = ffn_dim_multiplier
+        self.norm_eps = norm_eps
+        self.class_dropout_prob = class_dropout_prob
+        self.num_classes = num_classes
+        self.learn_sigma = learn_sigma
+        self.qk_norm = qk_norm
+
+        self.gradient_checkpointing = True
+        self.fused_attn = True
+
+        self.x_embedder = nn.Linear(in_channels * patch_size**2, dim)
+        self.t_embedder = TimestepEmbedder(min(dim, 1024))
+        self.y_embedder = LabelEmbedding(num_classes, min(dim, 1024), class_dropout_prob)
+
+        # 2. Define transformers blocks
+        self.layers = nn.LayerList(
+            [
+                TransformerBlock(
+                    layer_id=idx,
+                    dim=dim,
+                    n_heads=num_attention_heads,
+                    n_kv_heads=n_kv_heads,
+                    multiple_of=multiple_of,
+                    mlp_ratio=mlp_ratio,
+                    ffn_dim_multiplier=ffn_dim_multiplier,
+                    norm_eps=norm_eps,
+                    qk_norm=qk_norm,
+                    fused_attn=self.fused_attn,
+                )
+                for idx in range(num_layers)
+            ]
+        )
+
+        # 3. Define output layers
+        self.final_layer = FinalLayer(dim, patch_size, self.out_channels)
+        self.freqs_cis = self.precompute_freqs_cis(dim // num_attention_heads, 4096)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def enable_gradient_checkpointing(self, value=True):
+        self.gradient_checkpointing = value
+
+    def enable_xformers_memory_efficient_attention(self, attention_op: Optional[str] = None):
+        self._use_memory_efficient_attention_xformers = True
+        self.fused_attn = True
+
+    def unpatchify(self, x):
+        """
+        Args:
+            x: (N, T, patch_size**2 * C)
+            imgs: (N, H, W, C)
+        """
+        c = self.out_channels
+        p = self.patch_size
+        h = w = int(tuple(x.shape)[1] ** 0.5)
+        assert h * w == tuple(x.shape)[1]
+
+        x = x.reshape(shape=([tuple(x.shape)[0], h, w, p, p, c]))
+        x = paddle.einsum("nhwpqc->nchpwq", x)
+        imgs = x.reshape(shape=([tuple(x.shape)[0], c, h * p, h * p]))
+        return imgs
+
+    def patchify(self, x):
+        B, C, H, W = tuple(x.shape)
+        assert (H, W) == (self.sample_size, self.sample_size)
+        pH = pW = self.patch_size
+        x = x.reshape([B, C, H // pH, pH, W // pW, pW])
+        x = x.transpose([0, 2, 4, 1, 3, 5]).flatten(start_axis=-3).flatten(start_axis=1, stop_axis=2)
+        return x
+
+    @staticmethod
+    def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0):
+        """
+        Precompute the frequency tensor for complex exponentials (cis) with
+        given dimensions.
+
+        This function calculates a frequency tensor with complex exponentials
+        using the given dimension 'dim' and the end index 'end'. The 'theta'
+        parameter scales the frequencies. The returned tensor contains complex
+        values in complex64 data type.
+
+        Args:
+            dim (int): Dimension of the frequency tensor.
+            end (int): End index for precomputing frequencies.
+            theta (float, optional): Scaling factor for frequency computation.
+                Defaults to 10000.0.
+
+        Returns:
+            paddle.Tensor: Precomputed frequency tensor with complex
+                exponentials.
+        """
+        freqs = 1.0 / theta ** (paddle.arange(start=0, end=dim, step=2)[: dim // 2].cast("float32") / dim)
+        t = paddle.arange(end=end)
+        input_0, vec2_0 = TypePromote(t, freqs)
+        freqs = paddle.outer(input_0, vec2_0).cast("float32")
+        freqs_cis = paddle.complex(
+            paddle.ones_like(freqs) * paddle.cos(freqs), paddle.ones_like(freqs) * paddle.sin(freqs)
+        )
+        return freqs_cis
+
+    def forward(
+        self,
+        hidden_states: paddle.Tensor,
+        timestep: paddle.Tensor,
+        class_labels: paddle.Tensor,
+        return_dict: bool = True,
+    ):
+        """
+        Args:
+            hidden_states: (N, C, H, W) tensor of spatial inputs (images or latent
+                representations of images)
+            timestep: (N,) tensor of diffusion timesteps
+            class_labels: (N,) tensor of class labels
+        """
+        hidden_states = hidden_states.cast(self.dtype)
+        timestep = timestep.cast(self.dtype)
+
+        # 1. Input
+        hidden_states = self.patchify(hidden_states)
+        x = self.x_embedder(hidden_states)
+        t = self.t_embedder(timestep)
+        y = self.y_embedder(class_labels)
+        adaln_input = t + y
+
+        # 2. Blocks
+        for i, layer in enumerate(self.layers):
+            if self.gradient_checkpointing:
+                x = paddle.distributed.fleet.utils.recompute(layer, x, self.freqs_cis[: x.shape[1]], adaln_input)
+            else:
+                x = layer(
+                    x,
+                    self.freqs_cis[: x.shape[1]],
+                    adaln_input,
+                )
+
+        # 3. Output
+        hidden_states = self.final_layer(x, adaln_input)
+        output = self.unpatchify(hidden_states)
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/models/dual_transformer_2d.py

@@ -0,0 +1,158 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Optional
+
+import paddle.nn as nn
+
+from .transformer_2d import Transformer2DModel, Transformer2DModelOutput
+
+
+class DualTransformer2DModel(nn.Layer):
+    """
+    Dual transformer wrapper that combines two `Transformer2DModel`s for mixed inference.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            Pass if the input is continuous. The number of channels in the input and output.
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.1): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use.
+        sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images.
+            Note that this is fixed at training time as it is used for learning a number of position embeddings. See
+            `ImagePositionalEmbeddings`.
+        num_vector_embeds (`int`, *optional*):
+            Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels.
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`.
+            The number of diffusion steps used during training. Note that this is fixed at training time as it is used
+            to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for
+            up to but not more than steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the TransformerBlocks' attention should contain a bias parameter.
+    """
+
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: Optional[int] = None,
+        num_vector_embeds: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+    ):
+        super().__init__()
+        self.transformers = nn.LayerList(
+            [
+                Transformer2DModel(
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    in_channels=in_channels,
+                    num_layers=num_layers,
+                    dropout=dropout,
+                    norm_num_groups=norm_num_groups,
+                    cross_attention_dim=cross_attention_dim,
+                    attention_bias=attention_bias,
+                    sample_size=sample_size,
+                    num_vector_embeds=num_vector_embeds,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                )
+                for _ in range(2)
+            ]
+        )
+
+        # Variables that can be set by a pipeline:
+
+        # The ratio of transformer1 to transformer2's output states to be combined during inference
+        self.mix_ratio = 0.5
+
+        # The shape of `encoder_hidden_states` is expected to be
+        # `(batch_size, condition_lengths[0]+condition_lengths[1], num_features)`
+        self.condition_lengths = [77, 257]
+
+        # Which transformer to use to encode which condition.
+        # E.g. `(1, 0)` means that we'll use `transformers[1](conditions[0])` and `transformers[0](conditions[1])`
+        self.transformer_index_for_condition = [1, 0]
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states,
+        timestep=None,
+        added_cond_kwargs=None,
+        class_labels=None,
+        cross_attention_kwargs=None,
+        attention_mask=None,
+        encoder_attention_mask=None,
+        return_dict: bool = True,
+    ):
+        """
+        Args:
+            hidden_states ( When discrete, `paddle.Tensor` of shape `(batch size, num latent pixels)`.
+                When continuous, `paddle.Tensor` of shape `(batch size, channel, height, width)`): Input
+                hidden_states.
+            encoder_hidden_states ( `paddle.Tensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `paddle.Tensor`, *optional*):
+                Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step.
+            attention_mask (`paddle.Tensor`, *optional*):
+                Optional attention mask to be applied in Attention.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.transformer_2d.Transformer2DModelOutput`] or `tuple`:
+            [`~models.transformer_2d.Transformer2DModelOutput`] if `return_dict` is True, otherwise a `tuple`. When
+            returning a tuple, the first element is the sample tensor.
+        """
+        input_states = hidden_states
+
+        encoded_states = []
+        tokens_start = 0
+        # attention_mask is not used yet
+        for i in range(2):
+            # for each of the two transformers, pass the corresponding condition tokens
+            condition_state = encoder_hidden_states[:, tokens_start : tokens_start + self.condition_lengths[i]]
+            transformer_index = self.transformer_index_for_condition[i]
+            encoded_state = self.transformers[transformer_index](
+                input_states,
+                encoder_hidden_states=condition_state,
+                timestep=timestep,
+                cross_attention_kwargs=cross_attention_kwargs,
+                return_dict=False,
+            )[0]
+            encoded_states.append(encoded_state - input_states)
+            tokens_start += self.condition_lengths[i]
+
+        output_states = encoded_states[0] * self.mix_ratio + encoded_states[1] * (1 - self.mix_ratio)
+        output_states = output_states + input_states
+
+        if not return_dict:
+            return (output_states,)
+
+        return Transformer2DModelOutput(sample=output_states)

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/models/ema.py

@@ -0,0 +1,109 @@
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle
+from paddle import nn
+
+
+class LitEma(nn.Layer):
+    """
+    Exponential Moving Average (EMA) of model updates
+
+    Parameters:
+        model: The model architecture for apply EMA.
+        decay: The exponential decay. Default 0.9999.
+        use_num_updates: Whether to use number of updates when computing
+            averages.
+    """
+
+    def __init__(self, model, decay=0.9999, use_num_upates=True):
+        super().__init__()
+        if decay < 0.0 or decay > 1.0:
+            raise ValueError("Decay must be between 0 and 1")
+
+        self.m_name2s_name = {}
+        #  0-d tensor broadcasting is not supported during the Sharding initialization phase
+        # self.register_buffer("decay", paddle.to_tensor(decay, dtype=paddle.float32))
+        # self.register_buffer(
+        #     "num_updates",
+        #     paddle.to_tensor(0, dtype=paddle.int64) if use_num_upates else paddle.to_tensor(-1, dtype=paddle.int64),
+        # )
+        self.decay = decay
+        self.num_updates = 0 if use_num_upates else -1
+
+        for name, p in model.named_parameters():
+            if not p.stop_gradient:
+                # remove as '.'-character is not allowed in buffers
+                s_name = name.replace(".", "")
+                self.m_name2s_name.update({name: s_name})
+                self.register_buffer(s_name, p.clone().detach().astype('float32'))
+
+        self.collected_params = []
+
+    def forward(self, model):
+        decay = self.decay
+
+        if self.num_updates >= 0:
+            self.num_updates += 1
+            decay = min(self.decay, (1 + self.num_updates) / (10 + self.num_updates))
+
+        one_minus_decay = 1.0 - decay
+
+        with paddle.no_grad():
+            m_param = dict(model.named_parameters())
+            shadow_params = dict(self.named_buffers())
+
+            for key in m_param:
+                if not m_param[key].stop_gradient:
+                    sname = self.m_name2s_name[key]
+                    shadow_params[sname].scale_(decay)
+                    shadow_params[sname].add_(m_param[key].astype('float32') * one_minus_decay)
+                else:
+                    assert key not in self.m_name2s_name
+
+    def copy_to(self, model):
+        m_param = dict(model.named_parameters())
+        shadow_params = dict(self.named_buffers())
+        for key in m_param:
+            if not m_param[key].stop_gradient:
+                # allow dtype cast
+                m_param[key].copy_(shadow_params[self.m_name2s_name[key]].cast(m_param[key].dtype), False)
+            else:
+                assert key not in self.m_name2s_name
+
+    def store(self, parameters):
+        """
+        Save the current parameters for restoring later.
+        Args:
+          parameters: Iterable of `EagerParamBase`; the parameters to be
+            temporarily stored.
+        """
+        self.collected_params = [param.detach().cpu().clone() for param in parameters]
+
+    def restore(self, parameters):
+        """
+        Restore the parameters stored with the `store` method.
+        Useful to validate the model with EMA parameters without affecting the
+        original optimization process. Store the parameters before the
+        `copy_to` method. After validation (or model saving), use this to
+        restore the former parameters.
+        Args:
+          parameters: Iterable of `EagerParamBase`; the parameters to be
+            updated with the stored parameters.
+        """
+        for c_param, param in zip(self.collected_params, parameters):
+            # allow dtype cast
+            param.copy_(c_param.cast(param.dtype), False)
+        self.collected_params = None

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/models/lvdm_aemodules3d.py

@@ -0,0 +1,219 @@
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+
+import numpy as np
+import paddle
+
+
+def silu(x):
+    return x * paddle.nn.functional.sigmoid(x=x)
+
+
+class SiLU(paddle.nn.Layer):
+    def __init__(self):
+        super(SiLU, self).__init__()
+
+    def forward(self, x):
+        return silu(x)
+
+
+def hinge_d_loss(logits_real, logits_fake):
+    loss_real = paddle.mean(x=paddle.nn.functional.relu(x=1.0 - logits_real))
+    loss_fake = paddle.mean(x=paddle.nn.functional.relu(x=1.0 + logits_fake))
+    d_loss = 0.5 * (loss_real + loss_fake)
+    return d_loss
+
+
+def vanilla_d_loss(logits_real, logits_fake):
+    d_loss = 0.5 * (
+        paddle.mean(x=paddle.nn.functional.softplus(x=-logits_real))
+        + paddle.mean(x=paddle.nn.functional.softplus(x=logits_fake))
+    )
+    return d_loss
+
+
+def Normalize(in_channels, norm_type="group"):
+    assert norm_type in ["group", "batch"]
+    if norm_type == "group":
+        return paddle.nn.GroupNorm(
+            num_groups=32, num_channels=in_channels, epsilon=1e-06, weight_attr=None, bias_attr=None
+        )
+    elif norm_type == "batch":
+        return paddle.nn.SyncBatchNorm(in_channels)
+
+
+class ResBlock(paddle.nn.Layer):
+    def __init__(
+        self,
+        in_channels,
+        out_channels=None,
+        conv_shortcut=False,
+        dropout=0.0,
+        norm_type="group",
+        padding_type="replicate",
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+        self.norm1 = Normalize(in_channels, norm_type)
+        self.conv1 = SamePadConv3d(in_channels, out_channels, kernel_size=3, padding_type=padding_type)
+        self.dropout = paddle.nn.Dropout(p=dropout)
+        self.norm2 = Normalize(in_channels, norm_type)
+        self.conv2 = SamePadConv3d(out_channels, out_channels, kernel_size=3, padding_type=padding_type)
+        if self.in_channels != self.out_channels:
+            self.conv_shortcut = SamePadConv3d(in_channels, out_channels, kernel_size=3, padding_type=padding_type)
+
+    def forward(self, x):
+        h = x
+        h = self.norm1(h)
+        h = silu(h)
+        h = self.conv1(h)
+        h = self.norm2(h)
+        h = silu(h)
+        h = self.conv2(h)
+        if self.in_channels != self.out_channels:
+            x = self.conv_shortcut(x)
+        return x + h
+
+
+class SamePadConv3d(paddle.nn.Layer):
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1, bias=True, padding_type="replicate"):
+        super().__init__()
+        if isinstance(kernel_size, int):
+            kernel_size = (kernel_size,) * 3
+        if isinstance(stride, int):
+            stride = (stride,) * 3
+        total_pad = tuple([(k - s) for k, s in zip(kernel_size, stride)])
+        pad_input = []
+        for p in total_pad[::-1]:
+            pad_input.append((p // 2 + p % 2, p // 2))
+        pad_input = sum(pad_input, tuple())
+        self.pad_input = pad_input
+        self.padding_type = padding_type
+        self.conv = paddle.nn.Conv3D(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=0,
+            bias_attr=bias,
+        )
+        self.weight = self.conv.weight
+
+    def forward(self, x):
+        return self.conv(
+            paddle.nn.functional.pad(x=x, pad=self.pad_input, mode=self.padding_type, data_format="NCDHW")
+        )
+
+
+class SamePadConvTranspose3d(paddle.nn.Layer):
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1, bias=True, padding_type="replicate"):
+        super().__init__()
+        if isinstance(kernel_size, int):
+            kernel_size = (kernel_size,) * 3
+        if isinstance(stride, int):
+            stride = (stride,) * 3
+        total_pad = tuple([(k - s) for k, s in zip(kernel_size, stride)])
+        pad_input = []
+        for p in total_pad[::-1]:
+            pad_input.append((p // 2 + p % 2, p // 2))
+        pad_input = sum(pad_input, tuple())
+        self.pad_input = pad_input
+        self.padding_type = padding_type
+        self.convt = paddle.nn.Conv3DTranspose(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=tuple([(k - 1) for k in kernel_size]),
+            bias_attr=bias,
+        )
+
+    def forward(self, x):
+        return self.convt(
+            paddle.nn.functional.pad(x=x, pad=self.pad_input, mode=self.padding_type, data_format="NCDHW")
+        )
+
+
+class Encoder(paddle.nn.Layer):
+    def __init__(
+        self, n_hiddens, downsample, z_channels, double_z, image_channel=3, norm_type="group", padding_type="replicate"
+    ):
+        super().__init__()
+        n_times_downsample = np.array([int(math.log2(d)) for d in downsample])
+        self.conv_blocks = paddle.nn.LayerList()
+        max_ds = n_times_downsample.max()
+        self.conv_first = SamePadConv3d(image_channel, n_hiddens, kernel_size=3, padding_type=padding_type)
+        for i in range(max_ds):
+            block = paddle.nn.Layer()
+            in_channels = n_hiddens * 2**i
+            out_channels = n_hiddens * 2 ** (i + 1)
+            stride = tuple([(2 if d > 0 else 1) for d in n_times_downsample])
+            block.down = SamePadConv3d(in_channels, out_channels, 4, stride=stride, padding_type=padding_type)
+            block.res = ResBlock(out_channels, out_channels, norm_type=norm_type)
+            self.conv_blocks.append(block)
+            n_times_downsample -= 1
+        self.final_block = paddle.nn.Sequential(
+            Normalize(out_channels, norm_type),
+            SiLU(),
+            SamePadConv3d(
+                out_channels,
+                2 * z_channels if double_z else z_channels,
+                kernel_size=3,
+                stride=1,
+                padding_type=padding_type,
+            ),
+        )
+        self.out_channels = out_channels
+
+    def forward(self, x):
+        h = self.conv_first(x)
+        for block in self.conv_blocks:
+            h = block.down(h)
+            h = block.res(h)
+        h = self.final_block(h)
+        return h
+
+
+class Decoder(paddle.nn.Layer):
+    def __init__(self, n_hiddens, upsample, z_channels, image_channel, norm_type="group"):
+        super().__init__()
+        n_times_upsample = np.array([int(math.log2(d)) for d in upsample])
+        max_us = n_times_upsample.max()
+        in_channels = z_channels
+        self.conv_blocks = paddle.nn.LayerList()
+        for i in range(max_us):
+            block = paddle.nn.Layer()
+            in_channels = in_channels if i == 0 else n_hiddens * 2 ** (max_us - i + 1)
+            out_channels = n_hiddens * 2 ** (max_us - i)
+            us = tuple([(2 if d > 0 else 1) for d in n_times_upsample])
+            block.up = SamePadConvTranspose3d(in_channels, out_channels, 4, stride=us)
+            block.res1 = ResBlock(out_channels, out_channels, norm_type=norm_type)
+            block.res2 = ResBlock(out_channels, out_channels, norm_type=norm_type)
+            self.conv_blocks.append(block)
+            n_times_upsample -= 1
+        self.conv_out = SamePadConv3d(out_channels, image_channel, kernel_size=3)
+
+    def forward(self, x):
+        h = x
+        for i, block in enumerate(self.conv_blocks):
+            h = block.up(h)
+            h = block.res1(h)
+            h = block.res2(h)
+        h = self.conv_out(h)
+        return h

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/models/lvdm_distributions.py

@@ -0,0 +1,100 @@
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+import paddle
+
+
+class AbstractDistribution:
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters, deterministic=False):
+        self.parameters = parameters
+        self.mean, self.logvar = paddle.chunk(x=parameters, chunks=2, axis=1)
+        self.logvar = paddle.clip(x=self.logvar, min=-30.0, max=20.0)
+        self.deterministic = deterministic
+        self.std = paddle.exp(x=(0.5 * self.logvar).astype("float32"))
+        self.var = paddle.exp(x=self.logvar.astype("float32"))
+        if self.deterministic:
+            self.var = self.std = paddle.zeros_like(x=self.mean)
+
+    def sample(self, noise=None):
+        if noise is None:
+            noise = paddle.randn(shape=self.mean.shape)
+        x = self.mean + self.std * noise
+        return x
+
+    def kl(self, other=None):
+        if self.deterministic:
+            return paddle.to_tensor(data=[0.0], dtype="float32")
+        elif other is None:
+            return 0.5 * paddle.sum(x=paddle.pow(x=self.mean, y=2) + self.var - 1.0 - self.logvar, axis=[1, 2, 3])
+        else:
+            return 0.5 * paddle.sum(
+                x=paddle.pow(x=self.mean - other.mean, y=2) / other.var
+                + self.var / other.var
+                - 1.0
+                - self.logvar
+                + other.logvar,
+                axis=[1, 2, 3],
+            )
+
+    def nll(self, sample, dims=[1, 2, 3]):
+        if self.deterministic:
+            return paddle.to_tensor(data=[0.0], dtype="float32")
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * paddle.sum(x=logtwopi + self.logvar + paddle.pow(x=sample - self.mean, y=2) / self.var, axis=dims)
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, paddle.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+    logvar1, logvar2 = [(x if isinstance(x, paddle.Tensor) else paddle.to_tensor(data=x)) for x in (logvar1, logvar2)]
+    return 0.5 * (
+        -1.0
+        + logvar2
+        - logvar1
+        + paddle.exp(x=(logvar1 - logvar2).astype("float32"))
+        + (mean1 - mean2) ** 2 * paddle.exp(x=(-logvar2).astype("float32"))
+    )

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/models/lvdm_unet_3d.py

@@ -0,0 +1,713 @@
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from abc import abstractmethod
+from dataclasses import dataclass
+
+import paddle
+from einops import rearrange
+from paddle.distributed.fleet.utils import recompute
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from .lvdm_attention_temporal import SpatialTemporalTransformer, STAttentionBlock
+from .lvdm_util import (
+    avg_pool_nd,
+    conv_nd,
+    linear,
+    nonlinearity,
+    normalization,
+    timestep_embedding,
+    zero_module,
+)
+from .modeling_utils import ModelMixin
+
+
+@dataclass
+class LVDMUNet3DModelOutput(BaseOutput):
+    """
+    Args:
+        sample (`paddle.Tensor` of shape `(batch_size, num_frames, num_channels, height, width)`):
+            Hidden states conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: paddle.Tensor
+
+
+def convert_module_to_f16(x):
+    pass
+
+
+def convert_module_to_f32(x):
+    pass
+
+
+class TimestepBlock(paddle.nn.Layer):
+    """
+    Any module where forward() takes timestep embeddings as a second argument.
+    """
+
+    @abstractmethod
+    def forward(self, x, emb):
+        """
+        Apply the module to `x` given `emb` timestep embeddings.
+        """
+
+
+class TimestepEmbedSequential(paddle.nn.Sequential, TimestepBlock):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+
+    def forward(self, x, emb, context=None, **kwargs):
+        for layer in self:
+            if isinstance(layer, TimestepBlock):
+                x = layer(x, emb, **kwargs)
+            # elif isinstance(layer, STTransformerClass):
+            elif isinstance(layer, SpatialTemporalTransformer):
+                x = layer(x, context, **kwargs)
+            else:
+                x = layer(x)
+        return x
+
+
+class Upsample(paddle.nn.Layer):
+    """
+    An upsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, kernel_size_t=3, padding_t=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = conv_nd(
+                dims, self.channels, self.out_channels, (kernel_size_t, 3, 3), padding=(padding_t, 1, 1)
+            )
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            x = paddle.nn.functional.interpolate(
+                x=x, size=(x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest", data_format="NCDHW"
+            )
+        else:
+            x = paddle.nn.functional.interpolate(x=x, scale_factor=2, mode="nearest")
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(paddle.nn.Layer):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, kernel_size_t=3, padding_t=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = conv_nd(
+                dims, self.channels, self.out_channels, (kernel_size_t, 3, 3), stride=stride, padding=(padding_t, 1, 1)
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class ResBlock(TimestepBlock):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param use_checkpoint: if True, use gradient checkpointing on this module.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    """
+
+    def __init__(
+        self,
+        channels,
+        emb_channels,
+        dropout,
+        out_channels=None,
+        use_conv=False,
+        use_scale_shift_norm=False,
+        dims=2,
+        use_checkpoint=False,
+        up=False,
+        down=False,
+        kernel_size_t=3,
+        padding_t=1,
+        nonlinearity_type="silu",
+        **kwargs
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_checkpoint = use_checkpoint
+        self.use_scale_shift_norm = use_scale_shift_norm
+        self.nonlinearity_type = nonlinearity_type
+        self.in_layers = paddle.nn.Sequential(
+            normalization(channels),
+            nonlinearity(nonlinearity_type),
+            conv_nd(dims, channels, self.out_channels, (kernel_size_t, 3, 3), padding=(padding_t, 1, 1)),
+        )
+        self.updown = up or down
+        if up:
+            self.h_upd = Upsample(channels, False, dims, kernel_size_t=kernel_size_t, padding_t=padding_t)
+            self.x_upd = Upsample(channels, False, dims, kernel_size_t=kernel_size_t, padding_t=padding_t)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims, kernel_size_t=kernel_size_t, padding_t=padding_t)
+            self.x_upd = Downsample(channels, False, dims, kernel_size_t=kernel_size_t, padding_t=padding_t)
+        else:
+            self.h_upd = self.x_upd = paddle.nn.Identity()
+        self.emb_layers = paddle.nn.Sequential(
+            nonlinearity(nonlinearity_type),
+            linear(emb_channels, 2 * self.out_channels if use_scale_shift_norm else self.out_channels),
+        )
+        self.out_layers = paddle.nn.Sequential(
+            normalization(self.out_channels),
+            nonlinearity(nonlinearity_type),
+            paddle.nn.Dropout(p=dropout),
+            zero_module(
+                conv_nd(dims, self.out_channels, self.out_channels, (kernel_size_t, 3, 3), padding=(padding_t, 1, 1))
+            ),
+        )
+        if self.out_channels == channels:
+            self.skip_connection = paddle.nn.Identity()
+        elif use_conv:
+            self.skip_connection = conv_nd(
+                dims, channels, self.out_channels, (kernel_size_t, 3, 3), padding=(padding_t, 1, 1)
+            )
+        else:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
+
+    def forward(self, x, emb, **kwargs):
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+        :param x: an [N x C x ...] Tensor of features.
+        :param emb: an [N x emb_channels] Tensor of timestep embeddings.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        if self.use_checkpoint:
+            return recompute(self._forward, x, emb)
+        else:
+            return self._forward(x, emb)
+
+    def _forward(self, x, emb):
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+        emb_out = self.emb_layers(emb).astype(h.dtype)
+        if emb_out.dim() == 3:
+            emb_out = rearrange(emb_out, "b t c -> b c t")
+        while len(emb_out.shape) < h.dim():
+            emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm:
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            scale, shift = paddle.chunk(x=emb_out, chunks=2, axis=1)
+            h = out_norm(h) * (1 + scale) + shift
+            h = out_rest(h)
+        else:
+            h = h + emb_out
+            h = self.out_layers(h)
+        out = self.skip_connection(x) + h
+        return out
+
+
+# def make_spatialtemporal_transformer(module_name='attention_temporal',
+#     class_name='SpatialTemporalTransformer'):
+#     module = __import__(f'.lvdm_attention_temporal', fromlist=[
+#         class_name])
+#     global STTransformerClass
+#     STTransformerClass = getattr(module, class_name)
+#     return STTransformerClass
+
+
+def make_spatialtemporal_transformer(module_name="attention_temporal", class_name="SpatialTemporalTransformer"):
+    # Todo: Support loading more types of transformers
+    assert module_name == "attention_temporal" and class_name == "SpatialTemporalTransformer"
+    return SpatialTemporalTransformer
+
+
+class LVDMUNet3DModel(ModelMixin, ConfigMixin):
+    """
+    The full UNet model with attention and timestep embedding.
+    :param in_channels: channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param attention_resolutions: a collection of downsample rates at which
+        attention will take place. May be a set, list, or tuple.
+        For example, if this contains 4, then at 4x downsampling, attention
+        will be used.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param use_checkpoint: use gradient checkpointing to reduce memory usage.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=3,
+        num_classes=None,
+        use_checkpoint=False,
+        use_fp16=False,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        transformer_depth=1,
+        context_dim=None,
+        legacy=True,
+        kernel_size_t=1,
+        padding_t=1,
+        use_temporal_transformer=False,
+        temporal_length=None,
+        use_relative_position=False,
+        nonlinearity_type="silu",
+        ST_transformer_module="attention_temporal",
+        ST_transformer_class="SpatialTemporalTransformer",
+        **kwargs
+    ):
+        super().__init__()
+        if use_temporal_transformer:
+            assert (
+                context_dim is not None
+            ), "Fool!! You forgot to include the dimension of your cross-attention conditioning..."
+        if context_dim is not None:
+            assert (
+                use_temporal_transformer
+            ), "Fool!! You forgot to use the temporal transformer for your cross-attention conditioning..."
+            from omegaconf.listconfig import ListConfig
+
+            if type(context_dim) == ListConfig:
+                context_dim = list(context_dim)
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+        if num_heads == -1:
+            assert num_head_channels != -1, "Either num_heads or num_head_channels has to be set"
+        if num_head_channels == -1:
+            assert num_heads != -1, "Either num_heads or num_head_channels has to be set"
+        self.image_size = image_size
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        self.num_res_blocks = num_res_blocks
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.num_classes = num_classes
+        self.use_checkpoint = use_checkpoint
+        # Todo: support customted self.dtype
+        # self.dtype = 'float16' if use_fp16 else 'float32'
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.use_relative_position = use_relative_position
+        self.temporal_length = temporal_length
+        self.nonlinearity_type = nonlinearity_type
+        time_embed_dim = model_channels * 4
+        self.time_embed_dim = time_embed_dim
+        self.time_embed = paddle.nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nonlinearity(nonlinearity_type),
+            linear(time_embed_dim, time_embed_dim),
+        )
+        if self.num_classes is not None:
+            self.label_emb = paddle.nn.Embedding(num_classes, time_embed_dim)
+        STTransformerClass = make_spatialtemporal_transformer(
+            module_name=ST_transformer_module, class_name=ST_transformer_class
+        )
+        self.input_blocks = paddle.nn.LayerList(
+            sublayers=[
+                TimestepEmbedSequential(
+                    conv_nd(dims, in_channels, model_channels, (kernel_size_t, 3, 3), padding=(padding_t, 1, 1))
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for _ in range(num_res_blocks):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                        kernel_size_t=kernel_size_t,
+                        padding_t=padding_t,
+                        nonlinearity_type=nonlinearity_type,
+                        **kwargs,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        dim_head = ch // num_heads if use_temporal_transformer else num_head_channels
+                    layers.append(
+                        STAttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads,
+                            num_head_channels=dim_head,
+                            temporal_length=temporal_length,
+                            use_relative_position=use_relative_position,
+                        )
+                        if not use_temporal_transformer
+                        else STTransformerClass(
+                            ch,
+                            num_heads,
+                            dim_head,
+                            depth=transformer_depth,
+                            context_dim=context_dim,
+                            temporal_length=temporal_length,
+                            use_relative_position=use_relative_position,
+                            **kwargs,
+                        )
+                    )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                            kernel_size_t=kernel_size_t,
+                            padding_t=padding_t,
+                            nonlinearity_type=nonlinearity_type,
+                            **kwargs,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch,
+                            conv_resample,
+                            dims=dims,
+                            out_channels=out_ch,
+                            kernel_size_t=kernel_size_t,
+                            padding_t=padding_t,
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        if legacy:
+            dim_head = ch // num_heads if use_temporal_transformer else num_head_channels
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+                kernel_size_t=kernel_size_t,
+                padding_t=padding_t,
+                nonlinearity_type=nonlinearity_type,
+                **kwargs,
+            ),
+            STAttentionBlock(
+                ch,
+                use_checkpoint=use_checkpoint,
+                num_heads=num_heads,
+                num_head_channels=dim_head,
+                temporal_length=temporal_length,
+                use_relative_position=use_relative_position,
+            )
+            if not use_temporal_transformer
+            else STTransformerClass(
+                ch,
+                num_heads,
+                dim_head,
+                depth=transformer_depth,
+                context_dim=context_dim,
+                temporal_length=temporal_length,
+                use_relative_position=use_relative_position,
+                **kwargs,
+            ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+                kernel_size_t=kernel_size_t,
+                padding_t=padding_t,
+                nonlinearity_type=nonlinearity_type,
+                **kwargs,
+            ),
+        )
+        self._feature_size += ch
+        self.output_blocks = paddle.nn.LayerList(sublayers=[])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(num_res_blocks + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    ResBlock(
+                        ch + ich,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=model_channels * mult,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                        kernel_size_t=kernel_size_t,
+                        padding_t=padding_t,
+                        nonlinearity_type=nonlinearity_type,
+                        **kwargs,
+                    )
+                ]
+                ch = model_channels * mult
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        dim_head = ch // num_heads if use_temporal_transformer else num_head_channels
+                    layers.append(
+                        STAttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads,
+                            num_head_channels=dim_head,
+                            temporal_length=temporal_length,
+                            use_relative_position=use_relative_position,
+                        )
+                        if not use_temporal_transformer
+                        else STTransformerClass(
+                            ch,
+                            num_heads,
+                            dim_head,
+                            depth=transformer_depth,
+                            context_dim=context_dim,
+                            temporal_length=temporal_length,
+                            use_relative_position=use_relative_position,
+                            **kwargs,
+                        )
+                    )
+                if level and i == num_res_blocks:
+                    out_ch = ch
+                    layers.append(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True,
+                            kernel_size_t=kernel_size_t,
+                            padding_t=padding_t,
+                            nonlinearity_type=nonlinearity_type,
+                            **kwargs,
+                        )
+                        if resblock_updown
+                        else Upsample(
+                            ch,
+                            conv_resample,
+                            dims=dims,
+                            out_channels=out_ch,
+                            kernel_size_t=kernel_size_t,
+                            padding_t=padding_t,
+                        )
+                    )
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+        self.out = paddle.nn.Sequential(
+            normalization(ch),
+            nonlinearity(nonlinearity_type),
+            zero_module(conv_nd(dims, model_channels, out_channels, (kernel_size_t, 3, 3), padding=(padding_t, 1, 1))),
+        )
+
+    def convert_to_fp16(self):
+        """
+        Convert the torso of the model to float16.
+        """
+        self.input_blocks.apply(fn=convert_module_to_f16)
+        self.middle_block.apply(fn=convert_module_to_f16)
+        self.output_blocks.apply(fn=convert_module_to_f16)
+
+    def convert_to_fp32(self):
+        """
+        Convert the torso of the model to float32.
+        """
+        self.input_blocks.apply(fn=convert_module_to_f32)
+        self.middle_block.apply(fn=convert_module_to_f32)
+        self.output_blocks.apply(fn=convert_module_to_f32)
+
+    def forward(self, x, timesteps=None, time_emb_replace=None, context=None, y=None, **kwargs):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :param context: conditioning plugged in via crossattn
+        :param y: an [N] Tensor of labels, if class-conditional.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        # Fix 0D tensor bug
+        if timesteps.ndim == 0:
+            timesteps = timesteps.unsqueeze(0)
+        hs = []
+        if time_emb_replace is None:
+            t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
+            emb = self.time_embed(t_emb)
+        else:
+            emb = time_emb_replace
+        if y is not None:
+            assert y.shape == (x.shape[0],)
+            emb = emb + self.label_emb(y)
+        h = x.astype(self.dtype)
+        for module in self.input_blocks:
+            h = module(h, emb, context, **kwargs)
+            hs.append(h)
+        h = self.middle_block(h, emb, context, **kwargs)
+        for module in self.output_blocks:
+            h = paddle.concat(x=[h, hs.pop()], axis=1)
+            h = module(h, emb, context, **kwargs)
+        h = h.astype(x.dtype)
+        h = self.out(h)
+
+        return LVDMUNet3DModelOutput(sample=h)
+
+
+class FrameInterpPredUNet(LVDMUNet3DModel):
+    """
+    A Unet for unconditional generation, frame prediction and interpolation.
+    may need to input `mask` to indicate condition, as well as noise level `s` for condition augmentation.
+    """
+
+    def __init__(self, image_size, in_channels, cond_aug_mode=None, *args, **kwargs):
+        super().__init__(image_size, in_channels, *args, **kwargs)
+        if cond_aug_mode == "time_embed":
+            self.time_embed_cond = paddle.nn.Sequential(
+                linear(self.model_channels, self.time_embed_dim),
+                nonlinearity(self.nonlinearity_type),
+                linear(self.time_embed_dim, self.time_embed_dim),
+            )
+        elif cond_aug_mode == "learned_embed":
+            pass
+
+    def forward(self, x, timesteps, context=None, y=None, s=None, mask=None, **kwargs):
+        # Fix 0D tensor bug
+        if timesteps.ndim == 0:
+            timesteps = timesteps.unsqueeze(0)
+        if s is not None:
+            s_emb = timestep_embedding(s, self.model_channels, repeat_only=False)
+            s_emb = self.time_embed_cond(s_emb)
+            t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
+            emb = self.time_embed(t_emb)
+            assert emb.dim() == 2
+            mask_ = mask[:, :, :, (0), (0)]
+            t = mask.shape[2]
+            emb_mix = (
+                emb.unsqueeze(axis=2).tile(repeat_times=[1, 1, t]) * (1 - mask_)
+                + s_emb.unsqueeze(axis=2).tile(repeat_times=[1, 1, t]) * mask_
+            )
+            assert emb_mix.dim() == 3
+            emb_mix = rearrange(emb_mix, "b c t -> b t c")
+            time_emb_replace = emb_mix
+            timesteps = None
+        else:
+            time_emb_replace = None
+            timesteps = timesteps
+        return super().forward(x, timesteps, time_emb_replace=time_emb_replace, context=context, y=y, **kwargs)

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/models/modeling_outputs.py

@@ -0,0 +1,44 @@
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+
+from ..utils import BaseOutput
+
+
+@dataclass
+class AutoencoderKLOutput(BaseOutput):
+    """
+    Output of AutoencoderKL encoding method.
+
+    Args:
+        latent_dist (`DiagonalGaussianDistribution`):
+            Encoded outputs of `Encoder` represented as the mean and logvar of `DiagonalGaussianDistribution`.
+            `DiagonalGaussianDistribution` allows for sampling latents from the distribution.
+    """
+
+    latent_dist: "DiagonalGaussianDistribution"  # noqa: F821
+
+@dataclass
+class Transformer2DModelOutput(BaseOutput):
+    """
+    The output of [`Transformer2DModel`].
+
+    Args:
+        sample (`paddle.Tensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
+            distributions for the unnoised latent pixels.
+    """
+
+    sample: "paddle.Tensor"  # noqa: F821

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/models/modelscope_autoencoder_img2vid.py

@@ -0,0 +1,487 @@
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Union
+
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..loaders import FromOriginalVAEMixin
+from ..utils import BaseOutput, apply_forward_hook
+from .modeling_utils import ModelMixin
+from .vae import DecoderOutput, DiagonalGaussianDistribution
+
+
+@paddle.no_grad()
+def get_first_stage_encoding(encoder_posterior):
+    scale_factor = 0.18215
+    if isinstance(encoder_posterior, DiagonalGaussianDistribution):
+        z = encoder_posterior.sample()
+    elif isinstance(encoder_posterior, paddle.Tensor):
+        z = encoder_posterior
+    else:
+        raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
+    return scale_factor * z
+
+
+@dataclass
+class AutoencoderKLOutput(BaseOutput):
+    """
+    Output of AutoencoderKL encoding method.
+
+    Args:
+        latent_dist (`DiagonalGaussianDistribution`):
+            Encoded outputs of `Encoder` represented as the mean and logvar of `DiagonalGaussianDistribution`.
+            `DiagonalGaussianDistribution` allows for sampling latents from the distribution.
+    """
+
+    latent_dist: DiagonalGaussianDistribution
+
+
+def Normalize(in_channels, num_groups=32):
+    return paddle.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, epsilon=1e-6)
+
+
+def nonlinearity(x):
+    # swish
+    return x * F.sigmoid(x)
+
+
+class ResnetBlock(nn.Layer):
+    def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False, dropout, temb_channels=512):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = paddle.nn.Conv2D(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+        if temb_channels > 0:
+            self.temb_proj = paddle.nn.Linear(temb_channels, out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = paddle.nn.Dropout(dropout)
+        self.conv2 = paddle.nn.Conv2D(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = paddle.nn.Conv2D(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+            else:
+                self.nin_shortcut = paddle.nn.Conv2D(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None]
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x + h
+
+
+class AttnBlock(nn.Layer):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = paddle.nn.Conv2D(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.k = paddle.nn.Conv2D(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.v = paddle.nn.Conv2D(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.proj_out = paddle.nn.Conv2D(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b, c, h, w = q.shape
+        q = q.reshape([b, c, h * w])
+        q = q.transpose([0, 2, 1])
+        k = k.reshape([b, c, h * w])
+        w_ = paddle.bmm(q, k)
+        w_ = w_ * (int(c) ** (-0.5))
+        w_ = F.softmax(w_, axis=2)
+
+        # attend to values
+        v = v.reshape([b, c, h * w])
+        w_ = w_.transpose([0, 2, 1])
+        h_ = paddle.bmm(v, w_)
+        h_ = h_.reshape([b, c, h, w])
+
+        h_ = self.proj_out(h_)
+
+        return x + h_
+
+
+class Upsample(nn.Layer):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = paddle.nn.Conv2D(in_channels, in_channels, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, x):
+        x = F.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Layer):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = paddle.nn.Conv2D(in_channels, in_channels, kernel_size=3, stride=2, padding=0)
+
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0, 1, 0, 1)
+            x = F.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = F.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+
+class Encoder(nn.Layer):
+    def __init__(
+        self,
+        ch,
+        out_ch,
+        ch_mult,
+        num_res_blocks,
+        attn_resolutions,
+        in_channels,
+        resolution,
+        z_channels,
+        dropout=0.0,
+        resamp_with_conv=True,
+        double_z=True,
+        use_linear_attn=False,
+        attn_type="vanilla",
+        **ignore_kwargs
+    ):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        # downsampling
+        self.conv_in = paddle.nn.Conv2D(in_channels, self.ch, kernel_size=3, stride=1, padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,) + tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.LayerList()
+        for i_level in range(self.num_resolutions):
+            block = nn.LayerList()
+            attn = nn.LayerList()
+            block_in = ch * in_ch_mult[i_level]
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(AttnBlock(block_in))
+            down = nn.Layer()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Layer()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout
+        )
+        self.mid.attn_1 = AttnBlock(block_in)
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout
+        )
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = paddle.nn.Conv2D(
+            block_in, 2 * z_channels if double_z else z_channels, kernel_size=3, stride=1, padding=1
+        )
+
+    def forward(self, x):
+        # timestep embedding
+        temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions - 1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Layer):
+    def __init__(
+        self,
+        ch,
+        out_ch,
+        ch_mult,
+        num_res_blocks,
+        attn_resolutions,
+        in_channels,
+        resolution,
+        z_channels,
+        resamp_with_conv=True,
+        dropout=0.0,
+        give_pre_end=False,
+        tanh_out=False,
+        use_linear_attn=False,
+        attn_type="vanilla",
+        **ignorekwargs
+    ):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+
+        # compute block_in and curr_res at lowest res
+        block_in = ch * ch_mult[self.num_resolutions - 1]
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+
+        # z to block_in
+        self.conv_in = paddle.nn.Conv2D(z_channels, block_in, kernel_size=3, stride=1, padding=1)
+
+        # middle
+        self.mid = nn.Layer()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout
+        )
+        self.mid.attn_1 = AttnBlock(block_in)
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout
+        )
+
+        # upsampling
+        self.up = nn.LayerList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.LayerList()
+            attn = nn.LayerList()
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(AttnBlock(block_in))
+            up = nn.Layer()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            if len(self.up) == 0:
+                self.up.append(up)
+            else:
+                self.up.insert(0, up)
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = paddle.nn.Conv2D(block_in, out_ch, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, z):
+        self.last_z_shape = z.shape
+
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h, temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        if self.tanh_out:
+            h = paddle.tanh(h)
+        return h
+
+
+class AutoencoderKL_imgtovideo(ModelMixin, ConfigMixin, FromOriginalVAEMixin):
+    """img to video AutoencoderKL"""
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        ch,
+        out_ch,
+        in_channels,
+        resolution,
+        z_channels,
+        embed_dim,
+        attn_resolutions,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks=2,
+        dropout=0.0,
+        resamp_with_conv=True,
+        double_z=True,
+        use_linear_attn=False,
+        attn_type="vanilla",
+        pretrained=None,
+        ignore_keys=[],
+        image_key="image",
+        colorize_nlabels=None,
+        monitor=None,
+        ema_decay=None,
+        learn_logvar=False,
+        **kwargs
+    ):
+        super().__init__()
+        self.learn_logvar = learn_logvar
+        self.image_key = image_key
+        self.encoder = Encoder(
+            ch=ch,
+            out_ch=out_ch,
+            ch_mult=ch_mult,
+            num_res_blocks=num_res_blocks,
+            attn_resolutions=attn_resolutions,
+            in_channels=in_channels,
+            resolution=resolution,
+            z_channels=z_channels,
+            double_z=double_z,
+        )
+        self.decoder = Decoder(
+            ch=ch,
+            out_ch=out_ch,
+            ch_mult=ch_mult,
+            num_res_blocks=num_res_blocks,
+            attn_resolutions=attn_resolutions,
+            in_channels=in_channels,
+            resolution=resolution,
+            z_channels=z_channels,
+        )
+
+        self.quant_conv = paddle.nn.Conv2D(2 * z_channels, 2 * embed_dim, 1)
+        self.post_quant_conv = paddle.nn.Conv2D(embed_dim, z_channels, 1)
+        self.embed_dim = embed_dim
+        if colorize_nlabels is not None:
+            assert type(colorize_nlabels) == int
+            self.register_buffer("colorize", paddle.randn([3, colorize_nlabels, 1, 1]))
+        if monitor is not None:
+            self.monitor = monitor
+
+        self.use_ema = ema_decay is not None
+
+    @apply_forward_hook
+    def encode(self, x: paddle.Tensor, return_dict: bool = True) -> AutoencoderKLOutput:
+        h = self.encoder(x)
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+
+        if not return_dict:
+            return (posterior,)
+
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    @apply_forward_hook
+    def decode(self, z: paddle.Tensor, return_dict: bool = True) -> Union[DecoderOutput, paddle.Tensor]:
+        z = self.post_quant_conv(z)
+        decoded = self.decoder(z)
+
+        if not return_dict:
+            return (decoded,)
+
+        return DecoderOutput(sample=decoded)
+
+    def forward(self, input, sample_posterior=True, return_dict: bool = True) -> Union[DecoderOutput, paddle.Tensor]:
+        posterior = self.encode(input).latent_dist
+        if sample_posterior:
+            z = posterior.sample()
+        else:
+            z = posterior.mode()
+        dec = self.decode(z).sample
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)

VLMEvalKit_old/PaddleMIX/ppdiffusers/ppdiffusers/models/paddleinfer_runtime.py

@@ -0,0 +1,399 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import os
+import shutil
+from pathlib import Path
+from typing import Dict, Optional, Union
+
+import numpy as np
+
+from ..configuration_utils import FrozenDict
+from ..utils import (
+    DIFFUSERS_CACHE,
+    FROM_HF_HUB,
+    HF_HUB_OFFLINE,
+    PADDLE_INFER_MODEL_NAME,
+    PADDLE_INFER_WEIGHTS_NAME,
+    PPDIFFUSERS_CACHE,
+    _add_variant,
+    _get_model_file,
+    is_paddle_available,
+    logging,
+)
+from ..version import VERSION as __version__
+
+__all__ = ["PaddleInferRuntimeModel"]
+
+if is_paddle_available():
+    import paddle
+    import paddle.inference as paddle_infer
+
+logger = logging.get_logger(__name__)
+
+
+class PaddleInferRuntimeModel:
+    def __init__(self, model=None, config=None, **kwargs):
+        logger.info("ppdiffusers.PaddleInferRuntimeModel")
+        self.model = model
+        self.config = config
+        self.model_save_dir = kwargs.get("model_save_dir", None)
+        self.latest_model_name = kwargs.get("latest_model_name", None)
+        self.latest_params_name = kwargs.get("latest_params_name", None)
+        if self.latest_model_name is None:
+            self.latest_model_name = PADDLE_INFER_MODEL_NAME
+        if self.latest_params_name is None:
+            self.latest_params_name = PADDLE_INFER_WEIGHTS_NAME
+
+    def __call__(self, **kwargs):
+        kwargs.pop("output_shape", None)
+        kwargs.pop("infer_op", None)
+        inputs = {}
+        # for k, v in kwargs.items():
+        #     if k == "timestep":
+        #         if v.ndim == 0:
+        #             # fix 0D tensor error
+        #             v = v.reshape((1,))
+        #         # fix dtype error
+        #         v = v.astype("float32")
+        #     inputs[k] = v
+        # input_names = self.model.get_input_names()
+        # for i, name in enumerate(input_names):
+        #     input_tensor = self.model.get_input_handle(name)
+        #     if name not in inputs:
+        #         raise ValueError(f"Input {name} is not in the model.")
+        #     if isinstance(inputs[name], int):
+        #         inputs[name] = paddle.to_tensor(inputs[name])
+        #         if inputs[name].ndim == 0:  # fix 0D tensor error
+        #             inputs[name] = inputs[name].reshape((1,))
+        #             logger.warning(f"Input {name} is 0D tensor, reshape to (1,)")
+        #     # if not isinstance(input_tensor, paddle.Tensor):
+        #     #     input_tensor = paddle.to_tensor(input_tensor)
+        #     #     logger.warning(f"Input {name} is not paddle tensor, convert to paddle tensor")
+        #     #     if input_tensor.ndim == 0:  # fix 0D tensor error
+        #     #         input_tensor = input_tensor.reshape((1,))
+        #     input_tensor.reshape(inputs[name].shape)
+        #     input_tensor.copy_from_cpu(inputs[name].numpy())
+        # # do the inference
+        # self.model.run()
+        # results = []
+        # # get out data from output tensor
+        # output_names = self.model.get_output_names()
+        # for i, name in enumerate(output_names):
+        #     output_tensor = self.model.get_output_handle(name)
+        #     output_data = output_tensor.copy_to_cpu()
+        #     results.append(paddle.to_tensor(output_data))
+        # return results
+        for k, v in kwargs.items():
+            if isinstance(v, int):
+                v = paddle.to_tensor(v)
+            if k == "timestep" or k == "num_frames":
+                if v.ndim == 0:
+                    # fix 0D tensor error
+                    v = v.reshape((1,))
+                # fix dtype error
+                v = v.astype("float32")
+            if isinstance(v, np.ndarray):
+                v = paddle.to_tensor(v)
+            inputs[k] = v
+        input_list = []
+        input_names = self.model.get_input_names()
+        for i, name in enumerate(input_names):
+            if name not in inputs:
+                raise ValueError(f"Input {name} is not in the model.")
+            input_list.append(inputs[name])
+        # do the inference (zero copy)
+        self.model.run(input_list)
+        results = []
+        # get out data from output tensor
+        output_names = self.model.get_output_names()
+        for i, name in enumerate(output_names):
+            output_tensor = self.model.get_output_handle(name)
+            output_data = output_tensor.copy_to_cpu()
+            results.append(paddle.to_tensor(output_data))
+        return results
+
+    @staticmethod
+    def load_model(
+        model_path: Union[str, Path],
+        params_path: Union[str, Path] = None,
+        use_optim_cache: bool = False,
+        infer_config: Optional["paddle_infer.Congig"] = None,
+    ):
+        """
+        Loads an FastDeploy Inference Model with fastdeploy.RuntimeOption
+        Arguments:
+            model_path (`str` or `Path`):
+                Model path from which to load
+            params_path (`str` or `Path`):
+                Params path from which to load
+            use_optim_cache (`bool`, *optional*, defaults to `False`):
+                Whether to automatically load the optimized parameters from cache.(If the cache does not exist, it will be automatically generated)
+            runtime_options (fd.RuntimeOption, *optional*):
+                The RuntimeOption of fastdeploy to initialize the fastdeploy runtime. Default setting
+                the device to cpu and the backend to paddle inference
+        """
+        if infer_config is None:
+            infer_config = paddle_infer.Config()
+
+        if use_optim_cache:
+            # 首次运行，自动生成优化模型
+            params_dir = os.path.dirname(params_path)
+            optim_cache_dir = os.path.join(params_dir, "_optim_cache")
+            if not os.path.exists(optim_cache_dir):
+                os.makedirs(optim_cache_dir)
+                infer_config.switch_ir_optim(True)
+                infer_config.set_optim_cache_dir(optim_cache_dir)
+                infer_config.enable_save_optim_model(True)
+            else:
+                # 第二次运行，加载缓存的optim模型
+                infer_config.switch_ir_optim(False)
+                optimized_params_path = os.path.join(optim_cache_dir, "_optimized.pdiparams")
+                optimized_model_path = os.path.join(optim_cache_dir, "_optimized.pdmodel")
+                model_path = optimized_model_path
+                params_path = optimized_params_path
+
+        infer_config.set_prog_file(model_path)
+        infer_config.set_params_file(params_path)
+        return paddle_infer.create_predictor(infer_config)
+
+    def _save_pretrained(
+        self,
+        save_directory: Union[str, Path],
+        model_file_name: Optional[str] = None,
+        params_file_name: Optional[str] = None,
+        **kwargs
+    ):
+        """
+        Save a model and its configuration file to a directory, so that it can be re-loaded using the
+        [`~FastDeployRuntimeModel.from_pretrained`] class method. It will always save the
+        latest_model_name.
+        Arguments:
+            save_directory (`str` or `Path`):
+                Directory where to save the model file.
+            model_file_name(`str`, *optional*):
+                Overwrites the default model file name from `"inference.pdmodel"` to `model_file_name`. This allows you to save the
+                model with a different name.
+            params_file_name(`str`, *optional*):
+                Overwrites the default model file name from `"inference.pdiparams"` to `params_file_name`. This allows you to save the
+                model with a different name.
+        """
+        model_file_name = model_file_name if model_file_name is not None else PADDLE_INFER_MODEL_NAME
+        params_file_name = params_file_name if params_file_name is not None else PADDLE_INFER_WEIGHTS_NAME
+
+        src_model_path = self.model_save_dir.joinpath(self.latest_model_name)
+        dst_model_path = Path(save_directory).joinpath(model_file_name)
+
+        try:
+            shutil.copyfile(src_model_path, dst_model_path)
+        except shutil.SameFileError:
+            pass
+
+    def save_pretrained(
+        self,
+        save_directory: Union[str, os.PathLike],
+        **kwargs,
+    ):
+        """
+        Save a model to a directory, so that it can be re-loaded using the [`~FastDeployRuntimeModel.from_pretrained`] class
+        method.:
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to which to save. Will be created if it doesn't exist.
+        """
+        if os.path.isfile(save_directory):
+            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
+            return
+
+        os.makedirs(save_directory, exist_ok=True)
+
+        # saving model weights/files
+        self._save_pretrained(save_directory, **kwargs)
+
+    @classmethod
+    def _from_pretrained(
+        cls,
+        pretrained_model_name_or_path: Union[str, Path],
+        model_file_name: Optional[str] = None,
+        params_file_name: Optional[str] = None,
+        use_auth_token: Optional[Union[bool, str, None]] = None,
+        revision: Optional[str] = None,
+        subfolder: Optional[str] = None,
+        force_download: bool = False,
+        cache_dir: Optional[str] = None,
+        infer_config: Optional["paddle_infer.Config"] = None,
+        use_optim_cache: bool = False,
+        from_hf_hub: Optional[bool] = False,
+        proxies: Optional[Dict] = None,
+        resume_download: bool = False,
+        local_files_only: bool = False,
+        user_agent: Union[Dict, str, None] = None,
+        is_onnx_model: bool = False,
+        **kwargs,
+    ):
+        """
+        Load a model from a directory or the HF Hub.
+        Arguments:
+            pretrained_model_name_or_path (`str` or `Path`):
+                Directory from which to load
+            model_file_name (`str`):
+                Overwrites the default model file name from `"inference.pdmodel"` to `file_name`. This allows you to load
+                different model files from the same repository or directory.
+            params_file_name (`str`):
+                Overwrites the default params file name from `"inference.pdiparams"` to `file_name`. This allows you to load
+                different model files from the same repository or directory.
+            use_auth_token (`str` or `bool`):
+                Is needed to load models from a private or gated repository
+            revision (`str`):
+                Revision is the specific model version to use. It can be a branch name, a tag name, or a commit id
+            cache_dir (`Union[str, Path]`, *optional*):
+                Path to a directory in which a downloaded pretrained model configuration should be cached if the
+                standard cache should not be used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            runtime_options (`fastdeploy.RuntimeOption`, *optional*):
+                The RuntimeOption of fastdeploy.
+            subfolder (`str`, *optional*, defaults to `""`):
+                In case the relevant files are located inside a subfolder of the model repo (either remote in
+                huggingface.co or downloaded locally), you can specify the folder name here.
+            kwargs (`Dict`, *optional*):
+                kwargs will be passed to the model during initialization
+        """
+
+        model_file_name = model_file_name if model_file_name is not None else PADDLE_INFER_MODEL_NAME
+        params_file_name = params_file_name if params_file_name is not None else PADDLE_INFER_WEIGHTS_NAME
+        config = None
+
+        # load model from local directory
+        if os.path.isdir(pretrained_model_name_or_path):
+            model_path = os.path.join(pretrained_model_name_or_path, model_file_name)
+            params_path = os.path.join(pretrained_model_name_or_path, params_file_name)
+
+            model = PaddleInferRuntimeModel.load_model(
+                model_path,
+                params_path,
+                infer_config=infer_config,
+                use_optim_cache=use_optim_cache,
+            )
+            # 加载模型配置文件
+            config_path = os.path.join(pretrained_model_name_or_path, "config.json")
+            if os.path.exists(config_path):
+                with open(config_path, "r", encoding="utf-8") as f:
+                    config = json.load(f)
+                    config = FrozenDict(config)
+            kwargs["model_save_dir"] = Path(pretrained_model_name_or_path)
+
+        # load model from hub or paddle bos
+        else:
+            model_cache_path = _get_model_file(
+                pretrained_model_name_or_path=pretrained_model_name_or_path,
+                weights_name=model_file_name,
+                subfolder=subfolder,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                revision=revision,
+                from_hf_hub=from_hf_hub,
+                proxies=proxies,
+                resume_download=resume_download,
+                local_files_only=local_files_only,
+                use_auth_token=use_auth_token,
+                user_agent=user_agent,
+            )
+
+            params_cache_path = _get_model_file(
+                pretrained_model_name_or_path=pretrained_model_name_or_path,
+                weights_name=params_file_name,
+                subfolder=subfolder,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                revision=revision,
+                from_hf_hub=from_hf_hub,
+                proxies=proxies,
+                resume_download=resume_download,
+                local_files_only=local_files_only,
+                use_auth_token=use_auth_token,
+                user_agent=user_agent,
+            )
+            kwargs["latest_params_name"] = Path(params_cache_path).name
+            kwargs["model_save_dir"] = Path(model_cache_path).parent
+            kwargs["latest_model_name"] = Path(model_cache_path).name
+
+            model = PaddleInferRuntimeModel.load_model(
+                model_cache_path,
+                params_cache_path,
+                infer_config=infer_config,
+                use_optim_cache=use_optim_cache,
+            )
+            # 加载模型配置文件
+            config_path = os.path.join(kwargs["model_save_dir"], "config.json")
+            if os.path.exists(config_path):
+                with open(config_path, "r", encoding="utf-8") as f:
+                    config = json.load(f)
+                    config = FrozenDict(config)
+
+        return cls(model=model, config=config, **kwargs)
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        pretrained_model_name_or_path: Union[str, Path],
+        model_file_name: Optional[str] = None,
+        params_file_name: Optional[str] = None,
+        infer_configs: Optional["paddle_infer.Config"] = None,
+        use_optim_cache: bool = False,
+        **kwargs,
+    ):
+        from_hf_hub = kwargs.pop("from_hf_hub", FROM_HF_HUB)
+        cache_dir = (
+            kwargs.pop("cache_dir", DIFFUSERS_CACHE) if from_hf_hub else kwargs.pop("cache_dir", PPDIFFUSERS_CACHE)
+        )
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
+        revision = kwargs.pop("revision", None)
+        subfolder = kwargs.pop("subfolder", None)
+        variant = kwargs.pop("variant", None)
+
+        user_agent = {
+            "ppdiffusers": __version__,
+            "file_type": "model",
+            "framework": "paddleinfer",
+        }
+
+        return cls._from_pretrained(
+            pretrained_model_name_or_path=pretrained_model_name_or_path,
+            model_file_name=_add_variant(model_file_name, variant),
+            params_file_name=_add_variant(params_file_name, variant),
+            use_auth_token=use_auth_token,
+            revision=revision,
+            subfolder=subfolder,
+            force_download=force_download,
+            cache_dir=cache_dir,
+            infer_config=infer_configs,
+            use_optim_cache=use_optim_cache,
+            from_hf_hub=from_hf_hub,
+            proxies=proxies,
+            resume_download=resume_download,
+            local_files_only=local_files_only,
+            user_agent=user_agent,
+            **kwargs,
+        )
+
+    @property
+    def dtype(self) -> Union[str, paddle.dtype]:
+        return "float32"